diff --git a/.clang-format b/.clang-format
new file mode 100644
index 000000000..30863c27a
--- /dev/null
+++ b/.clang-format
@@ -0,0 +1,29 @@
+# This file is used by clang-format to autoformat paddle source code
+#
+# The clang-format is part of llvm toolchain.
+# It need to install llvm and clang to format source code style.
+#
+# The basic usage is,
+# clang-format -i -style=file PATH/TO/SOURCE/CODE
+#
+# The -style=file implicit use ".clang-format" file located in one of
+# parent directory.
+# The -i means inplace change.
+#
+# The document of clang-format is
+# http://clang.llvm.org/docs/ClangFormat.html
+# http://clang.llvm.org/docs/ClangFormatStyleOptions.html
+---
+Language: Cpp
+BasedOnStyle: Google
+IndentWidth: 2
+TabWidth: 2
+ContinuationIndentWidth: 4
+MaxEmptyLinesToKeep: 2
+AccessModifierOffset: -2 # The private/protected/public has no indent in class
+Standard: Cpp11
+AllowAllParametersOfDeclarationOnNextLine: true
+BinPackParameters: false
+BinPackArguments: false
+...
+
diff --git a/.clang_format.hook b/.clang_format.hook
new file mode 100755
index 000000000..4cbc972bb
--- /dev/null
+++ b/.clang_format.hook
@@ -0,0 +1,15 @@
+#!/usr/bin/env bash
+set -e
+
+readonly VERSION="3.9"
+
+version=$(clang-format -version)
+
+if ! [[ $version == *"$VERSION"* ]]; then
+ echo "clang-format version check failed."
+ echo "a version contains '$VERSION' is needed, but get '$version'"
+ echo "you can install the right version, and make an soft-link to '\$PATH' env"
+ exit -1
+fi
+
+clang-format $@
diff --git a/.gitignore b/.gitignore
new file mode 100644
index 000000000..dde3895fc
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1,2 @@
+.DS_Store
+*.pyc
diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml
new file mode 100644
index 000000000..ede1c53a4
--- /dev/null
+++ b/.pre-commit-config.yaml
@@ -0,0 +1,35 @@
+- repo: https://github.com/pre-commit/mirrors-yapf.git
+ sha: v0.16.0
+ hooks:
+ - id: yapf
+ files: \.py$
+- repo: https://github.com/pre-commit/pre-commit-hooks
+ sha: a11d9314b22d8f8c7556443875b731ef05965464
+ hooks:
+ - id: check-merge-conflict
+ - id: check-symlinks
+ - id: detect-private-key
+ files: (?!.*paddle)^.*$
+ - id: end-of-file-fixer
+ files: \.md$
+ - id: trailing-whitespace
+ files: \.md$
+- repo: https://github.com/Lucas-C/pre-commit-hooks
+ sha: v1.0.1
+ hooks:
+ - id: forbid-crlf
+ files: \.md$
+ - id: remove-crlf
+ files: \.md$
+ - id: forbid-tabs
+ files: \.md$
+ - id: remove-tabs
+ files: \.md$
+- repo: local
+ hooks:
+ - id: clang-format
+ name: clang-format
+ description: Format files with ClangFormat
+ entry: bash .clang_format.hook -i
+ language: system
+ files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx|cuh|proto)$
diff --git a/.style.yapf b/.style.yapf
new file mode 100644
index 000000000..4741fb4f3
--- /dev/null
+++ b/.style.yapf
@@ -0,0 +1,3 @@
+[style]
+based_on_style = pep8
+column_limit = 80
diff --git a/.travis.yml b/.travis.yml
new file mode 100644
index 000000000..75c2c1351
--- /dev/null
+++ b/.travis.yml
@@ -0,0 +1,34 @@
+language: cpp
+cache: ccache
+sudo: required
+dist: trusty
+services:
+ - docker
+os:
+ - linux
+env:
+ - JOB=PRE_COMMIT
+
+addons:
+ apt:
+ packages:
+ - git
+ - python
+ - python-pip
+ - python2.7-dev
+
+before_install:
+ - sudo pip install -U virtualenv pre-commit pip
+ - docker pull paddlepaddle/paddle:latest
+
+script:
+ - exit_code=0
+ - .travis/precommit.sh || exit_code=$(( exit_code | $? ))
+ - docker run -i --rm -v "$PWD:/py_unittest" paddlepaddle/paddle:latest /bin/bash -c
+ 'cd /py_unittest; sh .travis/unittest.sh' || exit_code=$(( exit_code | $? ))
+ exit $exit_code
+
+notifications:
+ email:
+ on_success: change
+ on_failure: always
diff --git a/.travis/precommit.sh b/.travis/precommit.sh
new file mode 100755
index 000000000..bcbfb2bb5
--- /dev/null
+++ b/.travis/precommit.sh
@@ -0,0 +1,21 @@
+#!/bin/bash
+function abort(){
+ echo "Your commit not fit PaddlePaddle code style" 1>&2
+ echo "Please use pre-commit scripts to auto-format your code" 1>&2
+ exit 1
+}
+
+trap 'abort' 0
+set -e
+cd `dirname $0`
+cd ..
+export PATH=/usr/bin:$PATH
+pre-commit install
+
+if ! pre-commit run -a ; then
+ ls -lh
+ git diff --exit-code
+ exit 1
+fi
+
+trap : 0
diff --git a/.travis/unittest.sh b/.travis/unittest.sh
new file mode 100755
index 000000000..f27dc481a
--- /dev/null
+++ b/.travis/unittest.sh
@@ -0,0 +1,29 @@
+#!/bin/bash
+
+abort(){
+ echo "Run unittest failed" 1>&2
+ echo "Please check your code" 1>&2
+ exit 1
+}
+
+unittest(){
+ cd $1 > /dev/null
+ if [ -f "setup.sh" ]; then
+ sh setup.sh
+ export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH
+ fi
+ if [ $? != 0 ]; then
+ exit 1
+ fi
+ find . -name 'tests' -type d -print0 | \
+ xargs -0 -I{} -n1 bash -c \
+ 'python -m unittest discover -v -s {}'
+ cd - > /dev/null
+}
+
+trap 'abort' 0
+set -e
+
+unittest .
+
+trap : 0
diff --git a/README.md b/README.md
new file mode 100644
index 000000000..0dcf8b602
--- /dev/null
+++ b/README.md
@@ -0,0 +1,523 @@
+# DeepSpeech2 on PaddlePaddle
+
+*DeepSpeech2 on PaddlePaddle* is an open-source implementation of end-to-end Automatic Speech Recognition (ASR) engine, based on [Baidu's Deep Speech 2 paper](http://proceedings.mlr.press/v48/amodei16.pdf), with [PaddlePaddle](https://github.com/PaddlePaddle/Paddle) platform. Our vision is to empower both industrial application and academic research on speech recognition, via an easy-to-use, efficient and scalable implementation, including training, inference & testing module, distributed [PaddleCloud](https://github.com/PaddlePaddle/cloud) training, and demo deployment. Besides, several pre-trained models for both English and Mandarin are also released.
+
+## Table of Contents
+- [Installation](#installation)
+- [Getting Started](#getting-started)
+- [Data Preparation](#data-preparation)
+- [Training a Model](#training-a-model)
+- [Data Augmentation Pipeline](#data-augmentation-pipeline)
+- [Inference and Evaluation](#inference-and-evaluation)
+- [Running in Docker Container](#running-in-docker-container)
+- [Distributed Cloud Training](#distributed-cloud-training)
+- [Hyper-parameters Tuning](#hyper-parameters-tuning)
+- [Training for Mandarin Language](#training-for-mandarin-language)
+- [Trying Live Demo with Your Own Voice](#trying-live-demo-with-your-own-voice)
+- [Released Models](#released-models)
+- [Experiments and Benchmarks](#experiments-and-benchmarks)
+- [Questions and Help](#questions-and-help)
+
+
+
+## Installation
+
+For this project was developed in PaddlePaddle V2 API, which is not maintained officially any more, we only support [running it in Docker container](#running-in-docker-container), instead of building environment from source code. And we are going to release the update to the latest Paddle Fluid API very soon, please keep an eye on this project.
+
+## Getting Started
+
+Several shell scripts provided in `./examples` will help us to quickly give it a try, for most major modules, including data preparation, model training, case inference and model evaluation, with a few public dataset (e.g. [LibriSpeech](http://www.openslr.org/12/), [Aishell](http://www.openslr.org/33)). Reading these examples will also help you to understand how to make it work with your own data.
+
+Some of the scripts in `./examples` are configured with 8 GPUs. If you don't have 8 GPUs available, please modify `CUDA_VISIBLE_DEVICES` and `--trainer_count`. If you don't have any GPU available, please set `--use_gpu` to False to use CPUs instead. Besides, if out-of-memory problem occurs, just reduce `--batch_size` to fit.
+
+Let's take a tiny sampled subset of [LibriSpeech dataset](http://www.openslr.org/12/) for instance.
+
+- Go to directory
+
+ ```bash
+ cd examples/tiny
+ ```
+
+ Notice that this is only a toy example with a tiny sampled subset of LibriSpeech. If you would like to try with the complete dataset (would take several days for training), please go to `examples/librispeech` instead.
+- Prepare the data
+
+ ```bash
+ sh run_data.sh
+ ```
+
+ `run_data.sh` will download dataset, generate manifests, collect normalizer's statistics and build vocabulary. Once the data preparation is done, you will find the data (only part of LibriSpeech) downloaded in `~/.cache/paddle/dataset/speech/libri` and the corresponding manifest files generated in `./data/tiny` as well as a mean stddev file and a vocabulary file. It has to be run for the very first time you run this dataset and is reusable for all further experiments.
+- Train your own ASR model
+
+ ```bash
+ sh run_train.sh
+ ```
+
+ `run_train.sh` will start a training job, with training logs printed to stdout and model checkpoint of every pass/epoch saved to `./checkpoints/tiny`. These checkpoints could be used for training resuming, inference, evaluation and deployment.
+- Case inference with an existing model
+
+ ```bash
+ sh run_infer.sh
+ ```
+
+ `run_infer.sh` will show us some speech-to-text decoding results for several (default: 10) samples with the trained model. The performance might not be good now as the current model is only trained with a toy subset of LibriSpeech. To see the results with a better model, you can download a well-trained (trained for several days, with the complete LibriSpeech) model and do the inference:
+
+ ```bash
+ sh run_infer_golden.sh
+ ```
+- Evaluate an existing model
+
+ ```bash
+ sh run_test.sh
+ ```
+
+ `run_test.sh` will evaluate the model with Word Error Rate (or Character Error Rate) measurement. Similarly, you can also download a well-trained model and test its performance:
+
+ ```bash
+ sh run_test_golden.sh
+ ```
+
+More detailed information are provided in the following sections. Wish you a happy journey with the *DeepSpeech2 on PaddlePaddle* ASR engine!
+
+
+## Data Preparation
+
+### Generate Manifest
+
+*DeepSpeech2 on PaddlePaddle* accepts a textual **manifest** file as its data set interface. A manifest file summarizes a set of speech data, with each line containing some meta data (e.g. filepath, transcription, duration) of one audio clip, in [JSON](http://www.json.org/) format, such as:
+
+```
+{"audio_filepath": "/home/work/.cache/paddle/Libri/134686/1089-134686-0001.flac", "duration": 3.275, "text": "stuff it into you his belly counselled him"}
+{"audio_filepath": "/home/work/.cache/paddle/Libri/134686/1089-134686-0007.flac", "duration": 4.275, "text": "a cold lucid indifference reigned in his soul"}
+```
+
+To use your custom data, you only need to generate such manifest files to summarize the dataset. Given such summarized manifests, training, inference and all other modules can be aware of where to access the audio files, as well as their meta data including the transcription labels.
+
+For how to generate such manifest files, please refer to `data/librispeech/librispeech.py`, which will download data and generate manifest files for LibriSpeech dataset.
+
+### Compute Mean & Stddev for Normalizer
+
+To perform z-score normalization (zero-mean, unit stddev) upon audio features, we have to estimate in advance the mean and standard deviation of the features, with some training samples:
+
+```bash
+python tools/compute_mean_std.py \
+--num_samples 2000 \
+--specgram_type linear \
+--manifest_paths data/librispeech/manifest.train \
+--output_path data/librispeech/mean_std.npz
+```
+
+It will compute the mean and standard deviation of power spectrum feature with 2000 random sampled audio clips listed in `data/librispeech/manifest.train` and save the results to `data/librispeech/mean_std.npz` for further usage.
+
+
+### Build Vocabulary
+
+A vocabulary of possible characters is required to convert the transcription into a list of token indices for training, and in decoding, to convert from a list of indices back to text again. Such a character-based vocabulary can be built with `tools/build_vocab.py`.
+
+```bash
+python tools/build_vocab.py \
+--count_threshold 0 \
+--vocab_path data/librispeech/eng_vocab.txt \
+--manifest_paths data/librispeech/manifest.train
+```
+
+It will write a vocabuary file `data/librispeeech/eng_vocab.txt` with all transcription text in `data/librispeech/manifest.train`, without vocabulary truncation (`--count_threshold 0`).
+
+### More Help
+
+For more help on arguments:
+
+```bash
+python data/librispeech/librispeech.py --help
+python tools/compute_mean_std.py --help
+python tools/build_vocab.py --help
+```
+
+## Training a model
+
+`train.py` is the main caller of the training module. Examples of usage are shown below.
+
+- Start training from scratch with 8 GPUs:
+
+ ```
+ CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train.py --trainer_count 8
+ ```
+
+- Start training from scratch with 16 CPUs:
+
+ ```
+ python train.py --use_gpu False --trainer_count 16
+ ```
+- Resume training from a checkpoint:
+
+ ```
+ CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
+ python train.py \
+ --init_model_path CHECKPOINT_PATH_TO_RESUME_FROM
+ ```
+
+For more help on arguments:
+
+```bash
+python train.py --help
+```
+or refer to `example/librispeech/run_train.sh`.
+
+## Data Augmentation Pipeline
+
+Data augmentation has often been a highly effective technique to boost the deep learning performance. We augment our speech data by synthesizing new audios with small random perturbation (label-invariant transformation) added upon raw audios. You don't have to do the syntheses on your own, as it is already embedded into the data provider and is done on the fly, randomly for each epoch during training.
+
+Six optional augmentation components are provided to be selected, configured and inserted into the processing pipeline.
+
+ - Volume Perturbation
+ - Speed Perturbation
+ - Shifting Perturbation
+ - Online Bayesian normalization
+ - Noise Perturbation (need background noise audio files)
+ - Impulse Response (need impulse audio files)
+
+In order to inform the trainer of what augmentation components are needed and what their processing orders are, it is required to prepare in advance an *augmentation configuration file* in [JSON](http://www.json.org/) format. For example:
+
+```
+[{
+ "type": "speed",
+ "params": {"min_speed_rate": 0.95,
+ "max_speed_rate": 1.05},
+ "prob": 0.6
+},
+{
+ "type": "shift",
+ "params": {"min_shift_ms": -5,
+ "max_shift_ms": 5},
+ "prob": 0.8
+}]
+```
+
+When the `--augment_conf_file` argument of `trainer.py` is set to the path of the above example configuration file, every audio clip in every epoch will be processed: with 60% of chance, it will first be speed perturbed with a uniformly random sampled speed-rate between 0.95 and 1.05, and then with 80% of chance it will be shifted in time with a random sampled offset between -5 ms and 5 ms. Finally this newly synthesized audio clip will be feed into the feature extractor for further training.
+
+For other configuration examples, please refer to `conf/augmenatation.config.example`.
+
+Be careful when utilizing the data augmentation technique, as improper augmentation will do harm to the training, due to the enlarged train-test gap.
+
+## Inference and Evaluation
+
+### Prepare Language Model
+
+A language model is required to improve the decoder's performance. We have prepared two language models (with lossy compression) for users to download and try. One is for English and the other is for Mandarin. Users can simply run this to download the preprared language models:
+
+```bash
+cd models/lm
+sh download_lm_en.sh
+sh download_lm_ch.sh
+```
+
+If you wish to train your own better language model, please refer to [KenLM](https://github.com/kpu/kenlm) for tutorials. Here we provide some tips to show how we preparing our English and Mandarin language models. You can take it as a reference when you train your own.
+
+#### English LM
+
+The English corpus is from the [Common Crawl Repository](http://commoncrawl.org) and you can download it from [statmt](http://data.statmt.org/ngrams/deduped_en). We use part en.00 to train our English language model. There are some preprocessing steps before training:
+
+ * Characters not in \[A-Za-z0-9\s'\] (\s represents whitespace characters) are removed and Arabic numbers are converted to English numbers like 1000 to one thousand.
+ * Repeated whitespace characters are squeezed to one and the beginning whitespace characters are removed. Notice that all transcriptions are lowercase, so all characters are converted to lowercase.
+ * Top 400,000 most frequent words are selected to build the vocabulary and the rest are replaced with 'UNKNOWNWORD'.
+
+Now the preprocessing is done and we get a clean corpus to train the language model. Our released language model are trained with agruments '-o 5 --prune 0 1 1 1 1'. '-o 5' means the max order of language model is 5. '--prune 0 1 1 1 1' represents count thresholds for each order and more specifically it will prune singletons for orders two and higher. To save disk storage we convert the arpa file to 'trie' binary file with arguments '-a 22 -q 8 -b 8'. '-a' represents the maximum number of leading bits of pointers in 'trie' to chop. '-q -b' are quantization parameters for probability and backoff.
+
+#### Mandarin LM
+
+Different from the English language model, Mandarin language model is character-based where each token is a Chinese character. We use internal corpus to train the released Mandarin language models. The corpus contain billions of tokens. The preprocessing has tiny difference from English language model and main steps include:
+
+ * The beginning and trailing whitespace characters are removed.
+ * English punctuations and Chinese punctuations are removed.
+ * A whitespace character between two tokens is inserted.
+
+Please notice that the released language models only contain Chinese simplified characters. After preprocessing done we can begin to train the language model. The key training arguments for small LM is '-o 5 --prune 0 1 2 4 4' and '-o 5' for large LM. Please refer above section for the meaning of each argument. We also convert the arpa file to binary file using default settings.
+
+### Speech-to-text Inference
+
+An inference module caller `infer.py` is provided to infer, decode and visualize speech-to-text results for several given audio clips. It might help to have an intuitive and qualitative evaluation of the ASR model's performance.
+
+- Inference with GPU:
+
+ ```bash
+ CUDA_VISIBLE_DEVICES=0 python infer.py --trainer_count 1
+ ```
+
+- Inference with CPUs:
+
+ ```bash
+ python infer.py --use_gpu False --trainer_count 12
+ ```
+
+We provide two types of CTC decoders: *CTC greedy decoder* and *CTC beam search decoder*. The *CTC greedy decoder* is an implementation of the simple best-path decoding algorithm, selecting at each timestep the most likely token, thus being greedy and locally optimal. The [*CTC beam search decoder*](https://arxiv.org/abs/1408.2873) otherwise utilizes a heuristic breadth-first graph search for reaching a near global optimality; it also requires a pre-trained KenLM language model for better scoring and ranking. The decoder type can be set with argument `--decoding_method`.
+
+For more help on arguments:
+
+```
+python infer.py --help
+```
+or refer to `example/librispeech/run_infer.sh`.
+
+### Evaluate a Model
+
+To evaluate a model's performance quantitatively, please run:
+
+- Evaluation with GPUs:
+
+ ```bash
+ CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python test.py --trainer_count 8
+ ```
+
+- Evaluation with CPUs:
+
+ ```bash
+ python test.py --use_gpu False --trainer_count 12
+ ```
+
+The error rate (default: word error rate; can be set with `--error_rate_type`) will be printed.
+
+For more help on arguments:
+
+```bash
+python test.py --help
+```
+or refer to `example/librispeech/run_test.sh`.
+
+## Hyper-parameters Tuning
+
+The hyper-parameters $\alpha$ (language model weight) and $\beta$ (word insertion weight) for the [*CTC beam search decoder*](https://arxiv.org/abs/1408.2873) often have a significant impact on the decoder's performance. It would be better to re-tune them on the validation set when the acoustic model is renewed.
+
+`tools/tune.py` performs a 2-D grid search over the hyper-parameter $\alpha$ and $\beta$. You must provide the range of $\alpha$ and $\beta$, as well as the number of their attempts.
+
+- Tuning with GPU:
+
+ ```bash
+ CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
+ python tools/tune.py \
+ --trainer_count 8 \
+ --alpha_from 1.0 \
+ --alpha_to 3.2 \
+ --num_alphas 45 \
+ --beta_from 0.1 \
+ --beta_to 0.45 \
+ --num_betas 8
+ ```
+
+- Tuning with CPU:
+
+ ```bash
+ python tools/tune.py --use_gpu False
+ ```
+ The grid search will print the WER (word error rate) or CER (character error rate) at each point in the hyper-parameters space, and draw the error surface optionally. A proper hyper-parameters range should include the global minima of the error surface for WER/CER, as illustrated in the following figure.
+
+
+
+
An example error surface for tuning on the dev-clean set of LibriSpeech
+
+
+Usually, as the figure shows, the variation of language model weight ($\alpha$) significantly affect the performance of CTC beam search decoder. And a better procedure is to first tune on serveral data batches (the number can be specified) to find out the proper range of hyper-parameters, then change to the whole validation set to carray out an accurate tuning.
+
+After tuning, you can reset $\alpha$ and $\beta$ in the inference and evaluation modules to see if they really help improve the ASR performance. For more help
+
+```bash
+python tune.py --help
+```
+or refer to `example/librispeech/run_tune.sh`.
+
+## Running in Docker Container
+
+Docker is an open source tool to build, ship, and run distributed applications in an isolated environment. A Docker image for this project has been provided in [hub.docker.com](https://hub.docker.com) with all the dependencies installed, including the pre-built PaddlePaddle, CTC decoders, and other necessary Python and third-party packages. This Docker image requires the support of NVIDIA GPU, so please make sure its availiability and the [nvidia-docker](https://github.com/NVIDIA/nvidia-docker) has been installed.
+
+Take several steps to launch the Docker image:
+
+- Download the Docker image
+
+```bash
+nvidia-docker pull paddlepaddle/deep_speech:latest-gpu
+```
+
+- Clone this repository
+
+```
+git clone https://github.com/PaddlePaddle/DeepSpeech.git
+```
+
+- Run the Docker image
+
+```bash
+sudo nvidia-docker run -it -v $(pwd)/DeepSpeech:/DeepSpeech paddlepaddle/deep_speech:latest-gpu /bin/bash
+```
+Now go back and start from the [Getting Started](#getting-started) section, you can execute training, inference and hyper-parameters tuning similarly in the Docker container.
+
+## Distributed Cloud Training
+
+We also provide a cloud training module for users to do the distributed cluster training on [PaddleCloud](https://github.com/PaddlePaddle/cloud), to achieve a much faster training speed with multiple machines. To start with this, please first install PaddleCloud client and register a PaddleCloud account, as described in [PaddleCloud Usage](https://github.com/PaddlePaddle/cloud/blob/develop/doc/usage_cn.md#%E4%B8%8B%E8%BD%BD%E5%B9%B6%E9%85%8D%E7%BD%AEpaddlecloud).
+
+Please take the following steps to submit a training job:
+
+- Go to directory:
+
+ ```bash
+ cd cloud
+ ```
+- Upload data:
+
+ Data must be uploaded to PaddleCloud filesystem to be accessed within a cloud job. `pcloud_upload_data.sh` helps do the data packing and uploading:
+
+ ```bash
+ sh pcloud_upload_data.sh
+ ```
+
+ Given input manifests, `pcloud_upload_data.sh` will:
+
+ - Extract the audio files listed in the input manifests.
+ - Pack them into a specified number of tar files.
+ - Upload these tar files to PaddleCloud filesystem.
+ - Create cloud manifests by replacing local filesystem paths with PaddleCloud filesystem paths. New manifests will be used to inform the cloud jobs of audio files' location and their meta information.
+
+ It should be done only once for the very first time to do the cloud training. Later, the data is kept persisitent on the cloud filesystem and reusable for further job submissions.
+
+ For argument details please refer to [Train DeepSpeech2 on PaddleCloud](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/cloud).
+
+ - Configure training arguments:
+
+ Configure the cloud job parameters in `pcloud_submit.sh` (e.g. `NUM_NODES`, `NUM_GPUS`, `CLOUD_TRAIN_DIR`, `JOB_NAME` etc.) and then configure other hyper-parameters for training in `pcloud_train.sh` (just as what you do for local training).
+
+ For argument details please refer to [Train DeepSpeech2 on PaddleCloud](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/cloud).
+
+ - Submit the job:
+
+ By running:
+
+ ```bash
+ sh pcloud_submit.sh
+ ```
+ a training job has been submitted to PaddleCloud, with the job name printed to the console.
+
+ - Get training logs
+
+ Run this to list all the jobs you have submitted, as well as their running status:
+
+ ```bash
+ paddlecloud get jobs
+ ```
+
+ Run this, the corresponding job's logs will be printed.
+ ```bash
+ paddlecloud logs -n 10000 $REPLACED_WITH_YOUR_ACTUAL_JOB_NAME
+ ```
+
+For more information about the usage of PaddleCloud, please refer to [PaddleCloud Usage](https://github.com/PaddlePaddle/cloud/blob/develop/doc/usage_cn.md#提交任务).
+
+For more information about the DeepSpeech2 training on PaddleCloud, please refer to
+[Train DeepSpeech2 on PaddleCloud](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/cloud).
+
+## Training for Mandarin Language
+
+The key steps of training for Mandarin language are same to that of English language and we have also provided an example for Mandarin training with Aishell in ```examples/aishell```. As mentioned above, please execute ```sh run_data.sh```, ```sh run_train.sh```, ```sh run_test.sh``` and ```sh run_infer.sh``` to do data preparation, training, testing and inference correspondingly. We have also prepared a pre-trained model (downloaded by ./models/aishell/download_model.sh) for users to try with ```sh run_infer_golden.sh``` and ```sh run_test_golden.sh```. Notice that, different from English LM, the Mandarin LM is character-based and please run ```tools/tune.py``` to find an optimal setting.
+
+## Trying Live Demo with Your Own Voice
+
+Until now, an ASR model is trained and tested qualitatively (`infer.py`) and quantitatively (`test.py`) with existing audio files. But it is not yet tested with your own speech. `deploy/demo_server.py` and `deploy/demo_client.py` helps quickly build up a real-time demo ASR engine with the trained model, enabling you to test and play around with the demo, with your own voice.
+
+To start the demo's server, please run this in one console:
+
+```bash
+CUDA_VISIBLE_DEVICES=0 \
+python deploy/demo_server.py \
+--trainer_count 1 \
+--host_ip localhost \
+--host_port 8086
+```
+
+For the machine (might not be the same machine) to run the demo's client, please do the following installation before moving on.
+
+For example, on MAC OS X:
+
+```bash
+brew install portaudio
+pip install pyaudio
+pip install pynput
+```
+
+Then to start the client, please run this in another console:
+
+```bash
+CUDA_VISIBLE_DEVICES=0 \
+python -u deploy/demo_client.py \
+--host_ip 'localhost' \
+--host_port 8086
+```
+
+Now, in the client console, press the `whitespace` key, hold, and start speaking. Until finishing your utterance, release the key to let the speech-to-text results shown in the console. To quit the client, just press `ESC` key.
+
+Notice that `deploy/demo_client.py` must be run on a machine with a microphone device, while `deploy/demo_server.py` could be run on one without any audio recording hardware, e.g. any remote server machine. Just be careful to set the `host_ip` and `host_port` argument with the actual accessible IP address and port, if the server and client are running with two separate machines. Nothing should be done if they are running on one single machine.
+
+Please also refer to `examples/mandarin/run_demo_server.sh`, which will first download a pre-trained Mandarin model (trained with 3000 hours of internal speech data) and then start the demo server with the model. With running `examples/mandarin/run_demo_client.sh`, you can speak Mandarin to test it. If you would like to try some other models, just update `--model_path` argument in the script.
+
+For more help on arguments:
+
+```bash
+python deploy/demo_server.py --help
+python deploy/demo_client.py --help
+```
+
+## Released Models
+
+#### Speech Model Released
+
+Language | Model Name | Training Data | Hours of Speech
+:-----------: | :------------: | :----------: | -------:
+English | [LibriSpeech Model](https://deepspeech.bj.bcebos.com/eng_models/librispeech_model.tar.gz) | [LibriSpeech Dataset](http://www.openslr.org/12/) | 960 h
+English | [BaiduEN8k Model](https://deepspeech.bj.bcebos.com/demo_models/baidu_en8k_model.tar.gz) | Baidu Internal English Dataset | 8628 h
+Mandarin | [Aishell Model](https://deepspeech.bj.bcebos.com/mandarin_models/aishell_model.tar.gz) | [Aishell Dataset](http://www.openslr.org/33/) | 151 h
+Mandarin | [BaiduCN1.2k Model](https://deepspeech.bj.bcebos.com/demo_models/baidu_cn1.2k_model.tar.gz) | Baidu Internal Mandarin Dataset | 1204 h
+
+#### Language Model Released
+
+Language Model | Training Data | Token-based | Size | Descriptions
+:-------------:| :------------:| :-----: | -----: | :-----------------
+[English LM](https://deepspeech.bj.bcebos.com/en_lm/common_crawl_00.prune01111.trie.klm) | [CommonCrawl(en.00)](http://web-language-models.s3-website-us-east-1.amazonaws.com/ngrams/en/deduped/en.00.deduped.xz) | Word-based | 8.3 GB | Pruned with 0 1 1 1 1;
About 1.85 billion n-grams;
'trie' binary with '-a 22 -q 8 -b 8'
+[Mandarin LM Small](https://deepspeech.bj.bcebos.com/zh_lm/zh_giga.no_cna_cmn.prune01244.klm) | Baidu Internal Corpus | Char-based | 2.8 GB | Pruned with 0 1 2 4 4;
About 0.13 billion n-grams;
'probing' binary with default settings
+[Mandarin LM Large](https://deepspeech.bj.bcebos.com/zh_lm/zhidao_giga.klm) | Baidu Internal Corpus | Char-based | 70.4 GB | No Pruning;
About 3.7 billion n-grams;
'probing' binary with default settings
+
+## Experiments and Benchmarks
+
+#### Benchmark Results for English Models (Word Error Rate)
+
+Test Set | LibriSpeech Model | BaiduEN8K Model
+:--------------------- | ---------------: | -------------------:
+LibriSpeech Test-Clean | 6.85 | 5.41
+LibriSpeech Test-Other | 21.18 | 13.85
+VoxForge American-Canadian | 12.12 | 7.13
+VoxForge Commonwealth | 19.82 | 14.93
+VoxForge European | 30.15 | 18.64
+VoxForge Indian | 53.73 | 25.51
+Baidu Internal Testset | 40.75 | 8.48
+
+For reproducing benchmark results on VoxForge data, we provide a script to download data and generate VoxForge dialect manifest files. Please go to ```data/voxforge``` and execute ```sh run_data.sh``` to get VoxForge dialect manifest files. Notice that VoxForge data may keep updating and the generated manifest files may have difference from those we evaluated on.
+
+#### Benchmark Results for Mandarin Model (Character Error Rate)
+
+Test Set | BaiduCN1.2k Model
+:--------------------- | -------------------:
+Baidu Internal Testset | 12.64
+
+#### Acceleration with Multi-GPUs
+
+We compare the training time with 1, 2, 4, 8, 16 Tesla K40m GPUs (with a subset of LibriSpeech samples whose audio durations are between 6.0 and 7.0 seconds). And it shows that a **near-linear** acceleration with multiple GPUs has been achieved. In the following figure, the time (in seconds) cost for training is printed on the blue bars.
+
+
+
+| # of GPU | Acceleration Rate |
+| -------- | --------------: |
+| 1 | 1.00 X |
+| 2 | 1.97 X |
+| 4 | 3.74 X |
+| 8 | 6.21 X |
+|16 | 10.70 X |
+
+`tools/profile.sh` provides such a profiling tool.
+
+## Questions and Help
+
+You are welcome to submit questions and bug reports in [Github Issues](https://github.com/PaddlePaddle/DeepSpeech/issues). You are also welcome to contribute to this project.
diff --git a/README_cn.md b/README_cn.md
new file mode 100644
index 000000000..06bee58bf
--- /dev/null
+++ b/README_cn.md
@@ -0,0 +1,527 @@
+# DeepSpeech2
+
+*DeepSpeech2* 是一个采用[PaddlePaddle](https://github.com/PaddlePaddle/Paddle)平台的端到端自动语音识别（ASR）引擎的开源项目，具体原理请参考这篇论文[Baidu's Deep Speech 2 paper](http://proceedings.mlr.press/v48/amodei16.pdf)。
+我们的愿景是为语音识别在工业应用和学术研究上，提供易于使用、高效和可扩展的工具，包括训练，推理，测试模块，以及分布式的[PaddleCloud](https://github.com/PaddlePaddle/cloud)训练和demo部署。同时，我们还将发布一些预训练好的英语和普通话模型。
+
+## 目录
+- [安装](#安装)
+- [开始](#开始)
+- [数据准备](#数据准备)
+- [训练模型](#训练模型)
+- [数据增强管道](#数据增强管道)
+- [推断和评价](#推断和评价)
+- [在Docker容器上运行](#在Docker容器上运行)
+- [分布式云训练](#分布式云训练)
+- [超参数调整](#超参数调整)
+- [训练汉语语言](#训练汉语语言)
+- [用自己的声音尝试现场演示](#用自己的声音尝试现场演示)
+- [发布模型](#发布模型)
+- [试验和基准](#试验和基准)
+- [问题和帮助](#问题和帮助)
+
+## 安装
+
+因该项目基于 PaddlePaddle V2 API 开发，其已不再被官方维护，目前我们仅支持 [在 Docker 容器中运行该项目](#在Docker容器上运行)，而不支持从源码构建环境。我们很快会将这个项目升级到最新的 Paddle Fluid API，请保持关注。
+
+## 开始
+
+`./examples`里的一些shell脚本将帮助我们在一些公开数据集(比如：[LibriSpeech](http://www.openslr.org/12/), [Aishell](http://www.openslr.org/33)) 进行快速尝试，包括了数据准备，模型训练，案例推断和模型评价。阅读这些例子将帮助你理解如何应用你的数据集。
+
+`./examples`目录中的一些脚本配置使用了8个GPU。如果你没有8个可用的GPU，请修改`CUDA_VISIBLE_DEVICES`和`--trainer_count`。如果你没有可用的GPU，请设置`--use_gpu`为False，这样程序会用CPU代替GPU。另外如果发生内存不足的问题，减小`--batch_size`即可。
+
+让我们先看看[LibriSpeech dataset](http://www.openslr.org/12/)小样本集的例子。
+
+- 转到目录
+
+ ```bash
+ cd examples/tiny
+ ```
+
+ 注意这仅仅是LibriSpeech一个小数据集的例子。如果你想尝试完整的数据集（可能需要花好几天来训练模型），请使用这个路径`examples/librispeech`。
+- 准备数据
+
+ ```bash
+ sh run_data.sh
+ ```
+
+ 运行`run_data.sh`脚本将会下载数据集，产出manifests文件，收集一些归一化需要的统计信息并建立词表。当数据准备完成之后，下载完的数据（仅有LibriSpeech一部分）在`~/.cache/paddle/dataset/speech/libri`中；其对应的manifest文件，均值标准差和词表文件在`./data/tiny`中。在第一次执行的时候一定要执行这个脚本，在接下来所有的实验中我们都会用到这个数据集。
+- 训练你自己的ASR模型
+
+ ```bash
+ sh run_train.sh
+ ```
+
+ `run_train.sh`将会启动训练任务，训练日志会打印到stdout，并且模型每个时期(epoch)的检查点都会保存到`./checkpoints/tiny`目录中。这些检查点可以用来恢复训练，推断，评价和部署。
+- 用已有的模型进行案例推断
+
+ ```bash
+ sh run_infer.sh
+ ```
+
+ `run_infer.sh`将会利用训完的模型展现一些（默认10个）样本语音到文本的解码结果。由于当前模型只使用了LibriSpeech一部分数据集训练，因此性能可能不会太好。为了看到更好模型上的表现，你可以下载一个已训练好的模型（用完整的LibriSpeech训练了好几天）来做推断。
+
+ ```bash
+ sh run_infer_golden.sh
+ ```
+- 评价一个已经存在的模型
+
+ ```bash
+ sh run_test.sh
+ ```
+
+ `run_test.sh`能够利用误字率（或字符错误率）来评价模型。类似的，你可以下载一个完全训练好的模型来测试它的性能：
+
+ ```bash
+ sh run_test_golden.sh
+ ```
+
+更多细节会在接下来的章节中阐述。祝你在*语音识别: DeepSpeech2*ASR引擎学习中过得愉快！
+
+
+## 数据准备
+
+### 生成Manifest
+
+*语音识别: DeepSpeech2*接受文本**manifest**文件作为数据接口。manifest文件包含了一系列语音数据，其中每一行代表一个json格式的音频元数据（比如文件路径，描述，时长）。具体格式如下：
+
+```
+{"audio_filepath": "/home/work/.cache/paddle/Libri/134686/1089-134686-0001.flac", "duration": 3.275, "text": "stuff it into you his belly counselled him"}
+{"audio_filepath": "/home/work/.cache/paddle/Libri/134686/1089-134686-0007.flac", "duration": 4.275, "text": "a cold lucid indifference reigned in his soul"}
+```
+
+如果你要使用自定义数据，你只需要按照以上格式生成自己的manifest文件即可。训练，推断以及其他所有模块都能够根据manifest文件获取到音频数据，包括他们的元数据。
+
+关于如何生成manifest文件，请参考`data/librispeech/librispeech.py`。该脚本将会下载LibriSpeech数据集并生成manifest文件。
+
+### 计算均值和标准差用于归一化
+
+为了对音频特征进行z-score归一化（零均值，单位标准差），我们必须预估一些训练样本特征的均值和标准差：
+
+```bash
+python tools/compute_mean_std.py \
+--num_samples 2000 \
+--specgram_type linear \
+--manifest_paths data/librispeech/manifest.train \
+--output_path data/librispeech/mean_std.npz
+```
+
+以上这段代码会计算在`data/librispeech/manifest.train`路径中，2000个随机采样音频剪辑的功率谱特征均值和标准差，并将结果保存在`data/librispeech/mean_std.npz`中，方便以后使用。
+
+### 建立词表
+
+转换录音为索引用于训练，解码，再将一系列索引转换为文本等操作需要一个可能会出现字符集合的词表。`tools/build_vocab.py`脚本将生成这种基于字符的词表。
+
+```bash
+python tools/build_vocab.py \
+--count_threshold 0 \
+--vocab_path data/librispeech/eng_vocab.txt \
+--manifest_paths data/librispeech/manifest.train
+```
+
+他将`data/librispeech/manifest.train`目录中的所有录音文本写入词表文件`data/librispeeech/eng_vocab.txt`，并且没有词汇截断(`--count_threshold 0`)。
+
+### 更多帮助
+
+获得更多帮助：
+
+```bash
+python data/librispeech/librispeech.py --help
+python tools/compute_mean_std.py --help
+python tools/build_vocab.py --help
+```
+
+## 训练模型
+
+`train.py`是训练模块的主要调用者。使用示例如下。
+
+- 开始使用8片GPU训练：
+
+ ```
+ CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train.py --trainer_count 8
+ ```
+
+- 开始使用16片GPU训练：
+
+ ```
+ python train.py --use_gpu False --trainer_count 16
+ ```
+
+- 从检查点恢复训练：
+
+ ```
+ CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
+ python train.py \
+ --init_model_path CHECKPOINT_PATH_TO_RESUME_FROM
+ ```
+
+获得更多帮助：
+
+```bash
+python train.py --help
+```
+或参考 `example/librispeech/run_train.sh`.
+
+## 数据增强管道
+
+数据增强是用来提升深度学习性能的非常有效的技术。我们通过在原始音频中添加小随机扰动（标签不变转换）获得新音频来增强我们的语音数据。你不必自己合成，因为数据增强已经嵌入到数据提供者中，能在训练模型时每个epoch中随机的合成音频。
+
+目前提供六个可选的增强组件供选择，配置并插入处理流水线。
+
+ - 音量扰动
+ - 速度扰动
+ - 移动扰动
+ - 在线贝叶斯归一化
+ - 噪声干扰（需要背景噪音的音频文件）
+ - 脉冲响应（需要脉冲音频文件）
+
+为了让训练模块知道需要哪些增强组件以及它们的处理顺序，我们需要事先准备一个[JSON](http://www.json.org/)格式的*扩展配置文件*。例如：
+
+```
+[{
+ "type": "speed",
+ "params": {"min_speed_rate": 0.95,
+ "max_speed_rate": 1.05},
+ "prob": 0.6
+},
+{
+ "type": "shift",
+ "params": {"min_shift_ms": -5,
+ "max_shift_ms": 5},
+ "prob": 0.8
+}]
+```
+
+当`trainer.py`的`--augment_conf_file`参数被设置为上述示例配置文件的路径时，每个epoch中的每个音频片段都将被处理。首先，均匀随机采样速率会有60％的概率在0.95和1.05之间对音频片段进行速度扰动。然后，音频片段有80％的概率在时间上被挪移，挪移偏差值是-5毫秒和5毫秒之间的随机采样。最后，这个新合成的音频片段将被传送给特征提取器，以用于接下来的训练。
+
+有关其他配置实例，请参考`conf/augmenatation.config.example`.
+
+使用数据增强技术时要小心，由于扩大了训练和测试集的差异，不恰当的增强会对训练模型不利。
+
+## 推断和评价
+
+### 准备语言模型
+
+提升解码器的性能需要准备语言模型。我们准备了两种语言模型（有损压缩）供用户下载和尝试。一个是英语模型，另一个是普通话模型。用户可以执行以下命令来下载已经训练好的语言模型：
+
+```bash
+cd models/lm
+sh download_lm_en.sh
+sh download_lm_ch.sh
+```
+
+如果你想训练自己更好的语言模型，请参考[KenLM](https://github.com/kpu/kenlm)获取教程。在这里，我们提供一些技巧来展示我们如何准备我们的英语和普通话模型。开始训练的时候，你可以参考这些技巧。
+
+
+#### 英语语言模型
+
+英语语料库来自[Common Crawl Repository](http://commoncrawl.org)，您可以从[statmt](http://data.statmt.org/ngrams/deduped_en)下载它。我们使用en.00部分来训练我们的英语语言模型。训练前有一些预处理步骤如下：
+
+ * 不在\[A-Za-z0-9\s'\]（\s表示空白字符）中的字符将被删除，阿拉伯数字被转换为英文数字，比如“1000”转换为one thousand。
+ * 重复的空白字符被压缩为一个，并且开始的空白字符将被删除。请注意，所有的录音都是小写字母，因此所有字符都转换为小写字母。
+ * 选择前40万个最常用的单词来建立词表，其余部分将被替换为“UNKNOWNWORD”。
+
+现在预处理完成了，我们得到一个干净的语料库来训练语言模型。我们发布的语言模型版本使用了参数“-o 5 --prune 0 1 1 1 1”来训练。“-o 5”表示语言模型的最大order为5。“--prune 0 1 1 1 1”表示每个order的计数阈值，更具体地说，它将第2个以及更高的order修剪为单个。为了节省磁盘存储空间，我们将使用参数“-a 22 -q 8 -b 8”将arpa文件转换为“trie”二进制文件。“-a”表示在“trie”中用于切分的指针的最高位数。“-q -b”是概率和退避的量化参数。
+
+#### 普通话语言模型
+
+与英语语言模型不同的是，普通话语言模型是基于字符的，其中每一位都是中文汉字。我们使用内部语料库来训练发布的汉语语言模型。该语料库包含数十亿汉字。预处理阶段与英语语言模型差别很小，主要步骤包括：
+
+ * 删除开始和结尾的空白字符。
+ * 删除英文标点和中文标点。
+ * 在两个字符之间插入空白字符。
+
+请注意，发布的语言模型只包含中文简体字。预处理完成后，我们开始训练语言模型。这个小的语言模型训练关键参数是“-o 5 --prune 0 1 2 4 4”，“-o 5”是针对大语言模型。请参考上面的部分了解每个参数的含义。我们还使用默认设置将arpa文件转换为二进制文件。
+
+### 语音到文本推断
+
+推断模块调用者为`infer.py`，可以用来推断，解码，以及给一些给定音频剪辑进行可视化语音到文本的结果。这有助于对ASR模型的性能进行直观和定性的评估。
+
+- GPU版本的推断：
+
+ ```bash
+ CUDA_VISIBLE_DEVICES=0 python infer.py --trainer_count 1
+ ```
+
+- CPU版本的推断：
+
+ ```bash
+ python infer.py --use_gpu False --trainer_count 12
+ ```
+
+我们提供两种类型的CTC解码器：*CTC贪心解码器*和*CTC波束搜索解码器*。*CTC贪心解码器*是简单的最佳路径解码算法的实现，在每个时间步选择最可能的字符，因此是贪心的并且是局部最优的。[*CTC波束搜索解码器*](https://arxiv.org/abs/1408.2873)另外使用了启发式广度优先图搜索以达到近似全局最优; 它也需要预先训练的KenLM语言模型以获得更好的评分和排名。解码器类型可以用参数`--decoding_method`设置。
+
+获得更多帮助：
+
+```
+python infer.py --help
+```
+或参考`example/librispeech/run_infer.sh`.
+
+### 评估模型
+
+要定量评估模型的性能，请运行：
+
+- 带GPU版本评估
+
+ ```bash
+ CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python test.py --trainer_count 8
+ ```
+
+- CPU版本评估
+
+ ```bash
+ python test.py --use_gpu False --trainer_count 12
+ ```
+
+错误率（默认：误字率;可以用--error_rate_type设置）将被打印出来。
+
+获得更多帮助：
+
+```bash
+python test.py --help
+```
+或参考`example/librispeech/run_test.sh`.
+
+## 超参数调整
+
+[*CTC波束搜索解码器*](https://arxiv.org/abs/1408.2873)的超参数$\alpha$（语言模型权重）和$\beta$（单词插入权重）对解码器的性能有非常显著的影响。当声学模型更新时，最好在验证集上重新调整它们。
+
+`tools/tune.py`会进行2维网格查找超参数$\alpha$和$\beta$。您必须提供$\alpha$和$\beta$的范围，以及尝试的次数。
+
+- 带GPU版的调整：
+
+ ```bash
+ CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
+ python tools/tune.py \
+ --trainer_count 8 \
+ --alpha_from 1.0 \
+ --alpha_to 3.2 \
+ --num_alphas 45 \
+ --beta_from 0.1 \
+ --beta_to 0.45 \
+ --num_betas 8
+ ```
+
+- CPU版的调整：
+
+ ```bash
+ python tools/tune.py --use_gpu False
+ ```
+网格搜索将会在超参数空间的每个点处打印出WER(误字率)或者CER(字符错误率)，并且可选择绘出误差曲面。合适的超参数范围应包括WER/CER误差表面的全局最小值，如下图所示。
+
+
+
+
调整LibriSpeech的dev-clean集合的误差曲面示例
+
+
+通常，如图所示，语言模型权重（$\alpha$）的变化显著影响CTC波束搜索解码器的性能。更好的方法是首先调整多批数据（可指定数量）以找出适当的超参数范围，然后更改为整个验证集以进行精确调整。
+
+调整之后，您可以在推理和评价模块中重置$\alpha$和$\beta$，以检查它们是否真的有助于提高ASR性能。更多帮助如下：
+
+```bash
+python tune.py --help
+```
+或参考`example/librispeech/run_tune.sh`.
+
+## 在Docker容器上运行
+
+Docker是一个开源工具，用于在孤立的环境中构建，发布和运行分布式应用程序。此项目的Docker镜像已在[hub.docker.com](https://hub.docker.com)中提供，并安装了所有依赖项，其中包括预先构建的PaddlePaddle，CTC解码器以及其他必要的Python和第三方库。这个Docker映像需要NVIDIA GPU的支持，所以请确保它的可用性并已完成[nvidia-docker](https://github.com/NVIDIA/nvidia-docker)的安装。
+
+采取以下步骤来启动Docker镜像：
+
+- 下载Docker镜像
+
+```bash
+nvidia-docker pull paddlepaddle/deep_speech:latest-gpu
+```
+
+- git clone这个资源库
+
+```
+git clone https://github.com/PaddlePaddle/DeepSpeech.git
+```
+
+- 运行Docker镜像
+
+```bash
+sudo nvidia-docker run -it -v $(pwd)/DeepSpeech:/DeepSpeech paddlepaddle/deep_speech:latest-gpu /bin/bash
+```
+
+现在返回并从[开始](#开始)部分开始，您可以在Docker容器中同样执行模型训练，推断和超参数调整。
+
+## 分布式云训练
+
+我们还为用户提供云训练模块[PaddleCloud](https://github.com/PaddlePaddle/cloud)以便用户进行集群训练，利用多台机器达到更快的训练速度。首先，请按照[PaddleCloud用法](https://github.com/PaddlePaddle/cloud/blob/develop/doc/usage_cn.md#%E4%B8%8B%E8%BD%BD%E5%B9%B6%E9%85%8D%E7%BD%AEpaddlecloud)安装PaddleCloud客户端并注册PaddleCloud账户。
+
+请按照以下步骤提交训练任务：
+
+- 转到目录：
+
+ ```bash
+ cd cloud
+ ```
+- 上传数据：
+
+ 数据必须上传到PaddleCloud文件系统才能在云作业中访问。`pcloud_upload_data.sh`负责进行数据打包和上传：
+
+ ```bash
+ sh pcloud_upload_data.sh
+ ```
+
+ 给定manifest文件，`pcloud_upload_data.sh`会进行以下处理：
+
+ - 提取输入清单中列出的音频文件。
+ - 将它们打包成指定数量的tar文件。
+ - 将这些tar文件上传到PaddleCloud文件系统。
+ - 通过用PaddleCloud文件系统路径替换本地文件系统路径来创建云manifest文件。云作业将通过新的manifest文件获取到音频文件的位置及其元信息。
+
+ 对于云训练模型来说以上步骤只需做一次。之后这些数据会在云文件系统上保持不变，并可在之后的任务中反复使用。
+
+ 有关参数的详细信息，请参考[在PaddleCloud上训练DeepSpeech2](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/cloud)。
+
+ - 配置训练参数
+
+ 在`pcloud_submit.sh`中配置云任务参数（例如`NUM_NODES`，`NUM_GPUS`，`CLOUD_TRAIN_DIR`，`JOB_NAME`等），然后在`pcloud_train.sh`中配置其他的超参数训练（和本地训练一样）。
+
+ 有关参数的详细信息，请参阅[在PaddleCloud上训练DeepSpeech2](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/cloud)。
+
+
+ - 提交任务
+
+ 运行：
+
+ ```bash
+ sh pcloud_submit.sh
+ ```
+ 一个训练任务已经提交给PaddleCloud，并将任务名输出到控制台。
+
+ - 获取训练日志
+
+ 执行以下命令以列出你提交的所有任务以及它们的运行状态：
+
+ ```bash
+ paddlecloud get jobs
+ ```
+
+ 运行此操作，将打印相应的任务日志。
+
+ ```bash
+ paddlecloud logs -n 10000 $REPLACED_WITH_YOUR_ACTUAL_JOB_NAME
+ ```
+
+有关PaddleCloud用法的更多信息，请参阅[PaddleCloud用法](https://github.com/PaddlePaddle/cloud/blob/develop/doc/usage_cn.md#提交任务)。
+
+有关PaddleCloud的DeepSpeech2训练的更多信息，请参阅
+[Train DeepSpeech2 on PaddleCloud](https://github.com/PaddlePaddle/DeepSpeech/tree/develop/cloud).
+
+## 训练普通话语言
+
+普通话语言训练与英语训练的关键步骤相同，我们提供了一个```examples/aishell```中Aishell的普通话训练例子。如上所述，请执行```sh run_data.sh```, ```sh run_train.sh```, ```sh run_test.sh```和```sh run_infer.sh```做相应的数据准备，训练，测试和推断。我们还准备了一个预训练过的模型（执行./models/aishell/download_model.sh下载）供用户使用```run_infer_golden.sh```和```run_test_golden.sh```来。请注意，与英语语言模型不同，普通话语言模型是基于汉字的，请运行```tools/tune.py```来查找最佳设置。
+
+##用自己的声音尝试现场演示
+
+到目前为止，一个ASR模型已经训练完毕，并且进行了定性测试（`infer.py`）和用现有的音频文件进行定量测试（`test.py`）。但目前还没有用你自己的声音进行测试。`deploy/demo_server.py`和`deploy/demo_client.py`能够快速构建一个利用训完的模型，对ASR引擎进行实时演示系统，使你能够用自己的语音测试和演示。
+
+要启动演示服务，请在控制台中运行：
+
+```bash
+CUDA_VISIBLE_DEVICES=0 \
+python deploy/demo_server.py \
+--trainer_count 1 \
+--host_ip localhost \
+--host_port 8086
+```
+
+对于运行demo客户端的机器（可能不是同一台机器），请在继续之前执行以下安装。
+
+比如，对于MAC OS X机器：
+
+```bash
+brew install portaudio
+pip install pyaudio
+pip install pynput
+```
+
+然后启动客户端，请在另一个控制台中运行：
+
+```bash
+CUDA_VISIBLE_DEVICES=0 \
+python -u deploy/demo_client.py \
+--host_ip 'localhost' \
+--host_port 8086
+```
+
+现在，在客户端控制台中，按下`whitespace`键，按住并开始讲话。讲话完毕请释放该键以让控制台中显示的语音到文本结果。要退出客户端，只需按`ESC`键。
+
+请注意，`deploy/demo_client.py`必须在带麦克风设备的机器上运行，而`deploy/demo_server.py`可以在没有任何录音硬件的情况下运行，例如任何远程服务器机器。如果服务器和客户端使用两台独立的机器运行，只需要注意将`host_ip`和`host_port`参数设置为实际可访问的IP地址和端口。如果它们在单台机器上运行，则不用作任何处理。
+
+请参考`examples/mandarin/run_demo_server.sh`，它将首先下载一个预先训练过的普通话模型（用3000小时的内部语音数据训练），然后用模型启动演示服务器。通过运行`examples/mandarin/run_demo_client.sh`，你可以说普通话来测试它。如果您想尝试其他模型，只需更新脚本中的`--model_path`参数即可。
+
+获得更多帮助：
+
+```bash
+python deploy/demo_server.py --help
+python deploy/demo_client.py --help
+```
+
+## 发布模型
+
+#### 语音模型发布
+
+语种 | 模型名 | 训练数据 | 语音时长
+:-----------: | :------------: | :----------: | -------:
+English | [LibriSpeech Model](https://deepspeech.bj.bcebos.com/eng_models/librispeech_model.tar.gz) | [LibriSpeech Dataset](http://www.openslr.org/12/) | 960 h
+English | [BaiduEN8k Model](https://deepspeech.bj.bcebos.com/demo_models/baidu_en8k_model.tar.gz) | Baidu Internal English Dataset | 8628 h
+Mandarin | [Aishell Model](https://deepspeech.bj.bcebos.com/mandarin_models/aishell_model.tar.gz) | [Aishell Dataset](http://www.openslr.org/33/) | 151 h
+Mandarin | [BaiduCN1.2k Model](https://deepspeech.bj.bcebos.com/demo_models/baidu_cn1.2k_model.tar.gz) | Baidu Internal Mandarin Dataset | 1204 h
+
+#### 语言模型发布
+
+语言模型 | 训练数据 | 基于的字符 | 大小 | 描述
+:-------------:| :------------:| :-----: | -----: | :-----------------
+[English LM](https://deepspeech.bj.bcebos.com/en_lm/common_crawl_00.prune01111.trie.klm) | [CommonCrawl(en.00)](http://web-language-models.s3-website-us-east-1.amazonaws.com/ngrams/en/deduped/en.00.deduped.xz) | Word-based | 8.3 GB | Pruned with 0 1 1 1 1;
About 1.85 billion n-grams;
'trie' binary with '-a 22 -q 8 -b 8'
+[Mandarin LM Small](https://deepspeech.bj.bcebos.com/zh_lm/zh_giga.no_cna_cmn.prune01244.klm) | Baidu Internal Corpus | Char-based | 2.8 GB | Pruned with 0 1 2 4 4;
About 0.13 billion n-grams;
'probing' binary with default settings
+[Mandarin LM Large](https://deepspeech.bj.bcebos.com/zh_lm/zhidao_giga.klm) | Baidu Internal Corpus | Char-based | 70.4 GB | No Pruning;
About 3.7 billion n-grams;
'probing' binary with default settings
+
+## 实验和基准
+
+#### 英语模型的基准测试结果（字错误率）
+
+测试集 | LibriSpeech Model | BaiduEN8K Model
+:--------------------- | ---------------: | -------------------:
+LibriSpeech Test-Clean | 6.85 | 5.41
+LibriSpeech Test-Other | 21.18 | 13.85
+VoxForge American-Canadian | 12.12 | 7.13
+VoxForge Commonwealth | 19.82 | 14.93
+VoxForge European | 30.15 | 18.64
+VoxForge Indian | 53.73 | 25.51
+Baidu Internal Testset | 40.75 | 8.48
+
+为了在VoxForge数据上重现基准测试结果，我们提供了一个脚本来下载数据并生成VoxForge方言manifest文件。请到```data/voxforge```执行````run_data.sh```来获取VoxForge方言manifest文件。请注意，VoxForge数据可能会持续更新，生成的清单文件可能与我们评估的清单文件有所不同。
+
+
+#### 普通话模型的基准测试结果（字符错误率）
+
+测试集 | BaiduCN1.2k Model
+:--------------------- | -------------------:
+Baidu Internal Testset | 12.64
+
+#### 多GPU加速
+
+我们对1,2,4,8,16个Tesla K40m GPU的训练时间（LibriSpeech样本的子集，其音频持续时间介于6.0和7.0秒之间）进行比较。它表明，已经实现了具有多个GPU的**近线性**加速。在下图中，训练的时间（以秒为单位）显示在蓝色条上。
+
+
+
+| # of GPU | 加速比 |
+| -------- | --------------: |
+| 1 | 1.00 X |
+| 2 | 1.97 X |
+| 4 | 3.74 X |
+| 8 | 6.21 X |
+|16 | 10.70 X |
+
+`tools/profile.sh`提供了上述分析工具.
+
+## 问题和帮助
+
+欢迎您在[Github问题](https://github.com/PaddlePaddle/models/issues)中提交问题和bug。也欢迎您为这个项目做出贡献。
diff --git a/cloud/README.md b/cloud/README.md
new file mode 100644
index 000000000..a5be1c420
--- /dev/null
+++ b/cloud/README.md
@@ -0,0 +1,63 @@
+# Train DeepSpeech2 on PaddleCloud
+
+>Note:
+>Please make sure [PaddleCloud Client](https://github.com/PaddlePaddle/cloud/blob/develop/doc/usage_cn.md#%E4%B8%8B%E8%BD%BD%E5%B9%B6%E9%85%8D%E7%BD%AEpaddlecloud) has be installed and current directory is `deep_speech_2/cloud/`
+
+## Step 1: Upload Data
+
+Provided with several input manifests, `pcloud_upload_data.sh` will pack and upload all the containing audio files to PaddleCloud filesystem, and also generate some corresponding manifest files with updated cloud paths.
+
+Please modify the following arguments in `pcloud_upload_data.sh`:
+
+- `IN_MANIFESTS`： Paths (in local filesystem) of manifest files containing the audio files to be uploaded. Multiple paths can be concatenated with a whitespace delimeter.
+- `OUT_MANIFESTS`: Paths (in local filesystem) to write the updated output manifest files to. Multiple paths can be concatenated with a whitespace delimeter. The values of `audio_filepath` in the output manifests are updated with cloud filesystem paths.
+- `CLOUD_DATA_DIR`: Directory (in PaddleCloud filesystem) to upload the data to. Don't forget to replace `USERNAME` in the default directory and make sure that you have the permission to write it.
+- `NUM_SHARDS`: Number of data shards / parts (in tar files) to be generated when packing and uploading data. Smaller `num_shards` requires larger temoporal local disk space for packing data.
+
+By running:
+
+```
+sh pcloud_upload_data.sh
+```
+all the audio files will be uploaded to PaddleCloud filesystem, and you will get modified manifests files in `OUT_MANIFESTS`.
+
+You have to take this step only once, in the very first time you do the cloud training. Later on, the data is persisitent on the cloud filesystem and reusable for further job submissions.
+
+## Step 2: Configure Training
+
+Configure cloud training arguments in `pcloud_submit.sh`, with the following arguments:
+
+- `TRAIN_MANIFEST`: Manifest filepath (in local filesystem) for training. Notice that the`audio_filepath` should be in cloud filesystem, like those generated by `pcloud_upload_data.sh`.
+- `DEV_MANIFEST`: Manifest filepath (in local filesystem) for validation.
+- `CLOUD_MODEL_DIR`: Directory (in PaddleCloud filesystem) to save the model parameters (checkpoints). Don't forget to replace `USERNAME` in the default directory and make sure that you have the permission to write it.
+- `BATCH_SIZE`: Training batch size for a single node.
+- `NUM_GPU`: Number of GPUs allocated for a single node.
+- `NUM_NODE`: Number of nodes (machines) allocated for this job.
+- `IS_LOCAL`: Set to False to enable parameter server, if using multiple nodes.
+
+Configure other training hyper-parameters in `pcloud_train.sh` as you wish, just as what you can do in local training.
+
+By running:
+
+```
+sh pcloud_submit.sh
+```
+you submit a training job to PaddleCloud. And you will see the job name when the submission is done.
+
+
+## Step 3 Get Job Logs
+
+Run this to list all the jobs you have submitted, as well as their running status:
+
+```
+paddlecloud get jobs
+```
+
+Run this, the corresponding job's logs will be printed.
+```
+paddlecloud logs -n 10000 $REPLACED_WITH_YOUR_ACTUAL_JOB_NAME
+```
+
+## More Help
+
+For more information about the usage of PaddleCloud, please refer to [PaddleCloud Usage](https://github.com/PaddlePaddle/cloud/blob/develop/doc/usage_cn.md#提交任务).
diff --git a/cloud/_init_paths.py b/cloud/_init_paths.py
new file mode 100644
index 000000000..3305d7488
--- /dev/null
+++ b/cloud/_init_paths.py
@@ -0,0 +1,17 @@
+"""Set up paths for DS2"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os.path
+import sys
+
+
+def add_path(path):
+ if path not in sys.path:
+ sys.path.insert(0, path)
+
+
+this_dir = os.path.dirname(__file__)
+proj_path = os.path.join(this_dir, '..')
+add_path(proj_path)
diff --git a/cloud/pcloud_submit.sh b/cloud/pcloud_submit.sh
new file mode 100644
index 000000000..99e458db9
--- /dev/null
+++ b/cloud/pcloud_submit.sh
@@ -0,0 +1,29 @@
+#! /usr/bin/env bash
+
+TRAIN_MANIFEST="cloud/cloud_manifests/cloud.manifest.train"
+DEV_MANIFEST="cloud/cloud_manifests/cloud.manifest.dev"
+CLOUD_MODEL_DIR="./checkpoints"
+BATCH_SIZE=512
+NUM_GPU=8
+NUM_NODE=1
+IS_LOCAL="True"
+
+JOB_NAME=deepspeech-`date +%Y%m%d%H%M%S`
+DS2_PATH=${PWD%/*}
+cp -f pcloud_train.sh ${DS2_PATH}
+
+paddlecloud submit \
+-image bootstrapper:5000/paddlepaddle/pcloud_ds2:latest \
+-jobname ${JOB_NAME} \
+-cpu ${NUM_GPU} \
+-gpu ${NUM_GPU} \
+-memory 64Gi \
+-parallelism ${NUM_NODE} \
+-pscpu 1 \
+-pservers 1 \
+-psmemory 64Gi \
+-passes 1 \
+-entry "sh pcloud_train.sh ${TRAIN_MANIFEST} ${DEV_MANIFEST} ${CLOUD_MODEL_DIR} ${NUM_GPU} ${BATCH_SIZE} ${IS_LOCAL}" \
+${DS2_PATH}
+
+rm ${DS2_PATH}/pcloud_train.sh
diff --git a/cloud/pcloud_train.sh b/cloud/pcloud_train.sh
new file mode 100644
index 000000000..d0c47dece
--- /dev/null
+++ b/cloud/pcloud_train.sh
@@ -0,0 +1,46 @@
+#! /usr/bin/env bash
+
+TRAIN_MANIFEST=$1
+DEV_MANIFEST=$2
+MODEL_PATH=$3
+NUM_GPU=$4
+BATCH_SIZE=$5
+IS_LOCAL=$6
+
+python ./cloud/split_data.py \
+--in_manifest_path=${TRAIN_MANIFEST} \
+--out_manifest_path='/local.manifest.train'
+
+python ./cloud/split_data.py \
+--in_manifest_path=${DEV_MANIFEST} \
+--out_manifest_path='/local.manifest.dev'
+
+mkdir ./logs
+
+python -u train.py \
+--batch_size=${BATCH_SIZE} \
+--trainer_count=${NUM_GPU} \
+--num_passes=200 \
+--num_proc_data=${NUM_GPU} \
+--num_conv_layers=2 \
+--num_rnn_layers=3 \
+--rnn_layer_size=2048 \
+--num_iter_print=100 \
+--learning_rate=5e-4 \
+--max_duration=27.0 \
+--min_duration=0.0 \
+--use_sortagrad=True \
+--use_gru=False \
+--use_gpu=True \
+--is_local=${IS_LOCAL} \
+--share_rnn_weights=True \
+--train_manifest='/local.manifest.train' \
+--dev_manifest='/local.manifest.dev' \
+--mean_std_path='data/librispeech/mean_std.npz' \
+--vocab_path='data/librispeech/vocab.txt' \
+--output_model_dir='./checkpoints' \
+--output_model_dir=${MODEL_PATH} \
+--augment_conf_path='conf/augmentation.config' \
+--specgram_type='linear' \
+--shuffle_method='batch_shuffle_clipped' \
+2>&1 | tee ./logs/train.log
diff --git a/cloud/pcloud_upload_data.sh b/cloud/pcloud_upload_data.sh
new file mode 100644
index 000000000..71bb4af19
--- /dev/null
+++ b/cloud/pcloud_upload_data.sh
@@ -0,0 +1,22 @@
+#! /usr/bin/env bash
+
+mkdir cloud_manifests
+
+IN_MANIFESTS="../data/librispeech/manifest.train ../data/librispeech/manifest.dev-clean ../data/librispeech/manifest.test-clean"
+OUT_MANIFESTS="cloud_manifests/cloud.manifest.train cloud_manifests/cloud.manifest.dev cloud_manifests/cloud.manifest.test"
+CLOUD_DATA_DIR="/pfs/dlnel/home/USERNAME/deepspeech2/data/librispeech"
+NUM_SHARDS=50
+
+python upload_data.py \
+--in_manifest_paths ${IN_MANIFESTS} \
+--out_manifest_paths ${OUT_MANIFESTS} \
+--cloud_data_dir ${CLOUD_DATA_DIR} \
+--num_shards ${NUM_SHARDS}
+
+if [ $? -ne 0 ]
+then
+ echo "Upload Data Failed!"
+ exit 1
+fi
+
+echo "All Done."
diff --git a/cloud/split_data.py b/cloud/split_data.py
new file mode 100644
index 000000000..3496d52bf
--- /dev/null
+++ b/cloud/split_data.py
@@ -0,0 +1,41 @@
+"""This tool is used for splitting data into each node of
+paddlecloud. This script should be called in paddlecloud.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import json
+import argparse
+
+parser = argparse.ArgumentParser(description=__doc__)
+parser.add_argument(
+ "--in_manifest_path",
+ type=str,
+ required=True,
+ help="Input manifest path for all nodes.")
+parser.add_argument(
+ "--out_manifest_path",
+ type=str,
+ required=True,
+ help="Output manifest file path for current node.")
+args = parser.parse_args()
+
+
+def split_data(in_manifest_path, out_manifest_path):
+ with open("/trainer_id", "r") as f:
+ trainer_id = int(f.readline()[:-1])
+ with open("/trainer_count", "r") as f:
+ trainer_count = int(f.readline()[:-1])
+
+ out_manifest = []
+ for index, json_line in enumerate(open(in_manifest_path, 'r')):
+ if (index % trainer_count) == trainer_id:
+ out_manifest.append("%s\n" % json_line.strip())
+ with open(out_manifest_path, 'w') as f:
+ f.writelines(out_manifest)
+
+
+if __name__ == '__main__':
+ split_data(args.in_manifest_path, args.out_manifest_path)
diff --git a/cloud/upload_data.py b/cloud/upload_data.py
new file mode 100644
index 000000000..9973f8c76
--- /dev/null
+++ b/cloud/upload_data.py
@@ -0,0 +1,129 @@
+"""This script is for uploading data for DeepSpeech2 training on paddlecloud.
+
+Steps:
+1. Read original manifests and extract local sound files.
+2. Tar all local sound files into multiple tar files and upload them.
+3. Modify original manifests with updated paths in cloud filesystem.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import json
+import os
+import tarfile
+import sys
+import argparse
+import shutil
+from subprocess import call
+import _init_paths
+from data_utils.utils import read_manifest
+
+parser = argparse.ArgumentParser(description=__doc__)
+parser.add_argument(
+ "--in_manifest_paths",
+ default=[
+ "../datasets/manifest.train", "../datasets/manifest.dev",
+ "../datasets/manifest.test"
+ ],
+ type=str,
+ nargs='+',
+ help="Local filepaths of input manifests to load, pack and upload."
+ "(default: %(default)s)")
+parser.add_argument(
+ "--out_manifest_paths",
+ default=[
+ "./cloud.manifest.train", "./cloud.manifest.dev",
+ "./cloud.manifest.test"
+ ],
+ type=str,
+ nargs='+',
+ help="Local filepaths of modified manifests to write to. "
+ "(default: %(default)s)")
+parser.add_argument(
+ "--cloud_data_dir",
+ required=True,
+ type=str,
+ help="Destination directory on paddlecloud to upload data to.")
+parser.add_argument(
+ "--num_shards",
+ default=10,
+ type=int,
+ help="Number of parts to split data to. (default: %(default)s)")
+parser.add_argument(
+ "--local_tmp_dir",
+ default="./tmp/",
+ type=str,
+ help="Local directory for storing temporary data. (default: %(default)s)")
+args = parser.parse_args()
+
+
+def upload_data(in_manifest_path_list, out_manifest_path_list, local_tmp_dir,
+ upload_tar_dir, num_shards):
+ """Extract and pack sound files listed in the manifest files into multple
+ tar files and upload them to padldecloud. Besides, generate new manifest
+ files with updated paths in paddlecloud.
+ """
+ # compute total audio number
+ total_line = 0
+ for manifest_path in in_manifest_path_list:
+ with open(manifest_path, 'r') as f:
+ total_line += len(f.readlines())
+ line_per_tar = (total_line // num_shards) + 1
+
+ # pack and upload shard by shard
+ line_count, tar_file = 0, None
+ for manifest_path, out_manifest_path in zip(in_manifest_path_list,
+ out_manifest_path_list):
+ manifest = read_manifest(manifest_path)
+ out_manifest = []
+ for json_data in manifest:
+ sound_filepath = json_data['audio_filepath']
+ sound_filename = os.path.basename(sound_filepath)
+ if line_count % line_per_tar == 0:
+ if tar_file != None:
+ tar_file.close()
+ pcloud_cp(tar_path, upload_tar_dir)
+ os.remove(tar_path)
+ tar_name = 'part-%s-of-%s.tar' % (
+ str(line_count // line_per_tar).zfill(5),
+ str(num_shards).zfill(5))
+ tar_path = os.path.join(local_tmp_dir, tar_name)
+ tar_file = tarfile.open(tar_path, 'w')
+ tar_file.add(sound_filepath, arcname=sound_filename)
+ line_count += 1
+ json_data['audio_filepath'] = "tar:%s#%s" % (
+ os.path.join(upload_tar_dir, tar_name), sound_filename)
+ out_manifest.append("%s\n" % json.dumps(json_data))
+ with open(out_manifest_path, 'w') as f:
+ f.writelines(out_manifest)
+ pcloud_cp(out_manifest_path, upload_tar_dir)
+ tar_file.close()
+ pcloud_cp(tar_path, upload_tar_dir)
+ os.remove(tar_path)
+
+
+def pcloud_mkdir(dir):
+ """Make directory in PaddleCloud filesystem.
+ """
+ if call(['paddlecloud', 'mkdir', dir]) != 0:
+ raise IOError("PaddleCloud mkdir failed: %s." % dir)
+
+
+def pcloud_cp(src, dst):
+ """Copy src from local filesytem to dst in PaddleCloud filesystem,
+ or downlowd src from PaddleCloud filesystem to dst in local filesystem.
+ """
+ if call(['paddlecloud', 'cp', src, dst]) != 0:
+ raise IOError("PaddleCloud cp failed: from [%s] to [%s]." % (src, dst))
+
+
+if __name__ == '__main__':
+ if not os.path.exists(args.local_tmp_dir):
+ os.makedirs(args.local_tmp_dir)
+ pcloud_mkdir(args.cloud_data_dir)
+
+ upload_data(args.in_manifest_paths, args.out_manifest_paths,
+ args.local_tmp_dir, args.cloud_data_dir, args.num_shards)
+
+ shutil.rmtree(args.local_tmp_dir)
diff --git a/conf/augmentation.config b/conf/augmentation.config
new file mode 100644
index 000000000..6c24da549
--- /dev/null
+++ b/conf/augmentation.config
@@ -0,0 +1,8 @@
+[
+ {
+ "type": "shift",
+ "params": {"min_shift_ms": -5,
+ "max_shift_ms": 5},
+ "prob": 1.0
+ }
+]
diff --git a/conf/augmentation.config.example b/conf/augmentation.config.example
new file mode 100644
index 000000000..21ed6ee10
--- /dev/null
+++ b/conf/augmentation.config.example
@@ -0,0 +1,39 @@
+[
+ {
+ "type": "noise",
+ "params": {"min_snr_dB": 40,
+ "max_snr_dB": 50,
+ "noise_manifest_path": "datasets/manifest.noise"},
+ "prob": 0.6
+ },
+ {
+ "type": "impulse",
+ "params": {"impulse_manifest_path": "datasets/manifest.impulse"},
+ "prob": 0.5
+ },
+ {
+ "type": "speed",
+ "params": {"min_speed_rate": 0.95,
+ "max_speed_rate": 1.05},
+ "prob": 0.5
+ },
+ {
+ "type": "shift",
+ "params": {"min_shift_ms": -5,
+ "max_shift_ms": 5},
+ "prob": 1.0
+ },
+ {
+ "type": "volume",
+ "params": {"min_gain_dBFS": -10,
+ "max_gain_dBFS": 10},
+ "prob": 0.0
+ },
+ {
+ "type": "bayesian_normal",
+ "params": {"target_db": -20,
+ "prior_db": -20,
+ "prior_samples": 100},
+ "prob": 0.0
+ }
+]
diff --git a/data/aishell/aishell.py b/data/aishell/aishell.py
new file mode 100644
index 000000000..d3340fc6f
--- /dev/null
+++ b/data/aishell/aishell.py
@@ -0,0 +1,110 @@
+"""Prepare Aishell mandarin dataset
+
+Download, unpack and create manifest files.
+Manifest file is a json-format file with each line containing the
+meta data (i.e. audio filepath, transcript and audio duration)
+of each audio file in the data set.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import codecs
+import soundfile
+import json
+import argparse
+from data_utils.utility import download, unpack
+
+DATA_HOME = os.path.expanduser('~/.cache/paddle/dataset/speech')
+
+URL_ROOT = 'http://www.openslr.org/resources/33'
+DATA_URL = URL_ROOT + '/data_aishell.tgz'
+MD5_DATA = '2f494334227864a8a8fec932999db9d8'
+
+parser = argparse.ArgumentParser(description=__doc__)
+parser.add_argument(
+ "--target_dir",
+ default=DATA_HOME + "/Aishell",
+ type=str,
+ help="Directory to save the dataset. (default: %(default)s)")
+parser.add_argument(
+ "--manifest_prefix",
+ default="manifest",
+ type=str,
+ help="Filepath prefix for output manifests. (default: %(default)s)")
+args = parser.parse_args()
+
+
+def create_manifest(data_dir, manifest_path_prefix):
+ print("Creating manifest %s ..." % manifest_path_prefix)
+ json_lines = []
+ transcript_path = os.path.join(data_dir, 'transcript',
+ 'aishell_transcript_v0.8.txt')
+ transcript_dict = {}
+ for line in codecs.open(transcript_path, 'r', 'utf-8'):
+ line = line.strip()
+ if line == '': continue
+ audio_id, text = line.split(' ', 1)
+ # remove withespace
+ text = ''.join(text.split())
+ transcript_dict[audio_id] = text
+
+ data_types = ['train', 'dev', 'test']
+ for type in data_types:
+ del json_lines[:]
+ audio_dir = os.path.join(data_dir, 'wav', type)
+ for subfolder, _, filelist in sorted(os.walk(audio_dir)):
+ for fname in filelist:
+ audio_path = os.path.join(subfolder, fname)
+ audio_id = fname[:-4]
+ # if no transcription for audio then skipped
+ if audio_id not in transcript_dict:
+ continue
+ audio_data, samplerate = soundfile.read(audio_path)
+ duration = float(len(audio_data) / samplerate)
+ text = transcript_dict[audio_id]
+ json_lines.append(
+ json.dumps(
+ {
+ 'audio_filepath': audio_path,
+ 'duration': duration,
+ 'text': text
+ },
+ ensure_ascii=False))
+ manifest_path = manifest_path_prefix + '.' + type
+ with codecs.open(manifest_path, 'w', 'utf-8') as fout:
+ for line in json_lines:
+ fout.write(line + '\n')
+
+
+def prepare_dataset(url, md5sum, target_dir, manifest_path):
+ """Download, unpack and create manifest file."""
+ data_dir = os.path.join(target_dir, 'data_aishell')
+ if not os.path.exists(data_dir):
+ filepath = download(url, md5sum, target_dir)
+ unpack(filepath, target_dir)
+ # unpack all audio tar files
+ audio_dir = os.path.join(data_dir, 'wav')
+ for subfolder, _, filelist in sorted(os.walk(audio_dir)):
+ for ftar in filelist:
+ unpack(os.path.join(subfolder, ftar), subfolder, True)
+ else:
+ print("Skip downloading and unpacking. Data already exists in %s." %
+ target_dir)
+ create_manifest(data_dir, manifest_path)
+
+
+def main():
+ if args.target_dir.startswith('~'):
+ args.target_dir = os.path.expanduser(args.target_dir)
+
+ prepare_dataset(
+ url=DATA_URL,
+ md5sum=MD5_DATA,
+ target_dir=args.target_dir,
+ manifest_path=args.manifest_prefix)
+
+
+if __name__ == '__main__':
+ main()
diff --git a/data/librispeech/librispeech.py b/data/librispeech/librispeech.py
new file mode 100644
index 000000000..9a8e1c287
--- /dev/null
+++ b/data/librispeech/librispeech.py
@@ -0,0 +1,148 @@
+"""Prepare Librispeech ASR datasets.
+
+Download, unpack and create manifest files.
+Manifest file is a json-format file with each line containing the
+meta data (i.e. audio filepath, transcript and audio duration)
+of each audio file in the data set.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import distutils.util
+import os
+import sys
+import argparse
+import soundfile
+import json
+import codecs
+from data_utils.utility import download, unpack
+
+URL_ROOT = "http://www.openslr.org/resources/12"
+URL_TEST_CLEAN = URL_ROOT + "/test-clean.tar.gz"
+URL_TEST_OTHER = URL_ROOT + "/test-other.tar.gz"
+URL_DEV_CLEAN = URL_ROOT + "/dev-clean.tar.gz"
+URL_DEV_OTHER = URL_ROOT + "/dev-other.tar.gz"
+URL_TRAIN_CLEAN_100 = URL_ROOT + "/train-clean-100.tar.gz"
+URL_TRAIN_CLEAN_360 = URL_ROOT + "/train-clean-360.tar.gz"
+URL_TRAIN_OTHER_500 = URL_ROOT + "/train-other-500.tar.gz"
+
+MD5_TEST_CLEAN = "32fa31d27d2e1cad72775fee3f4849a9"
+MD5_TEST_OTHER = "fb5a50374b501bb3bac4815ee91d3135"
+MD5_DEV_CLEAN = "42e2234ba48799c1f50f24a7926300a1"
+MD5_DEV_OTHER = "c8d0bcc9cca99d4f8b62fcc847357931"
+MD5_TRAIN_CLEAN_100 = "2a93770f6d5c6c964bc36631d331a522"
+MD5_TRAIN_CLEAN_360 = "c0e676e450a7ff2f54aeade5171606fa"
+MD5_TRAIN_OTHER_500 = "d1a0fd59409feb2c614ce4d30c387708"
+
+parser = argparse.ArgumentParser(description=__doc__)
+parser.add_argument(
+ "--target_dir",
+ default='~/.cache/paddle/dataset/speech/libri',
+ type=str,
+ help="Directory to save the dataset. (default: %(default)s)")
+parser.add_argument(
+ "--manifest_prefix",
+ default="manifest",
+ type=str,
+ help="Filepath prefix for output manifests. (default: %(default)s)")
+parser.add_argument(
+ "--full_download",
+ default="True",
+ type=distutils.util.strtobool,
+ help="Download all datasets for Librispeech."
+ " If False, only download a minimal requirement (test-clean, dev-clean"
+ " train-clean-100). (default: %(default)s)")
+args = parser.parse_args()
+
+
+def create_manifest(data_dir, manifest_path):
+ """Create a manifest json file summarizing the data set, with each line
+ containing the meta data (i.e. audio filepath, transcription text, audio
+ duration) of each audio file within the data set.
+ """
+ print("Creating manifest %s ..." % manifest_path)
+ json_lines = []
+ for subfolder, _, filelist in sorted(os.walk(data_dir)):
+ text_filelist = [
+ filename for filename in filelist if filename.endswith('trans.txt')
+ ]
+ if len(text_filelist) > 0:
+ text_filepath = os.path.join(data_dir, subfolder, text_filelist[0])
+ for line in open(text_filepath):
+ segments = line.strip().split()
+ text = ' '.join(segments[1:]).lower()
+ audio_filepath = os.path.join(data_dir, subfolder,
+ segments[0] + '.flac')
+ audio_data, samplerate = soundfile.read(audio_filepath)
+ duration = float(len(audio_data)) / samplerate
+ json_lines.append(
+ json.dumps({
+ 'audio_filepath': audio_filepath,
+ 'duration': duration,
+ 'text': text
+ }))
+ with codecs.open(manifest_path, 'w', 'utf-8') as out_file:
+ for line in json_lines:
+ out_file.write(line + '\n')
+
+
+def prepare_dataset(url, md5sum, target_dir, manifest_path):
+ """Download, unpack and create summmary manifest file.
+ """
+ if not os.path.exists(os.path.join(target_dir, "LibriSpeech")):
+ # download
+ filepath = download(url, md5sum, target_dir)
+ # unpack
+ unpack(filepath, target_dir)
+ else:
+ print("Skip downloading and unpacking. Data already exists in %s." %
+ target_dir)
+ # create manifest json file
+ create_manifest(target_dir, manifest_path)
+
+
+def main():
+ if args.target_dir.startswith('~'):
+ args.target_dir = os.path.expanduser(args.target_dir)
+
+ prepare_dataset(
+ url=URL_TEST_CLEAN,
+ md5sum=MD5_TEST_CLEAN,
+ target_dir=os.path.join(args.target_dir, "test-clean"),
+ manifest_path=args.manifest_prefix + ".test-clean")
+ prepare_dataset(
+ url=URL_DEV_CLEAN,
+ md5sum=MD5_DEV_CLEAN,
+ target_dir=os.path.join(args.target_dir, "dev-clean"),
+ manifest_path=args.manifest_prefix + ".dev-clean")
+ if args.full_download:
+ prepare_dataset(
+ url=URL_TRAIN_CLEAN_100,
+ md5sum=MD5_TRAIN_CLEAN_100,
+ target_dir=os.path.join(args.target_dir, "train-clean-100"),
+ manifest_path=args.manifest_prefix + ".train-clean-100")
+ prepare_dataset(
+ url=URL_TEST_OTHER,
+ md5sum=MD5_TEST_OTHER,
+ target_dir=os.path.join(args.target_dir, "test-other"),
+ manifest_path=args.manifest_prefix + ".test-other")
+ prepare_dataset(
+ url=URL_DEV_OTHER,
+ md5sum=MD5_DEV_OTHER,
+ target_dir=os.path.join(args.target_dir, "dev-other"),
+ manifest_path=args.manifest_prefix + ".dev-other")
+ prepare_dataset(
+ url=URL_TRAIN_CLEAN_360,
+ md5sum=MD5_TRAIN_CLEAN_360,
+ target_dir=os.path.join(args.target_dir, "train-clean-360"),
+ manifest_path=args.manifest_prefix + ".train-clean-360")
+ prepare_dataset(
+ url=URL_TRAIN_OTHER_500,
+ md5sum=MD5_TRAIN_OTHER_500,
+ target_dir=os.path.join(args.target_dir, "train-other-500"),
+ manifest_path=args.manifest_prefix + ".train-other-500")
+
+
+if __name__ == '__main__':
+ main()
diff --git a/data/noise/chime3_background.py b/data/noise/chime3_background.py
new file mode 100644
index 000000000..f79ca7335
--- /dev/null
+++ b/data/noise/chime3_background.py
@@ -0,0 +1,128 @@
+"""Prepare CHiME3 background data.
+
+Download, unpack and create manifest files.
+Manifest file is a json-format file with each line containing the
+meta data (i.e. audio filepath, transcript and audio duration)
+of each audio file in the data set.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import distutils.util
+import os
+import wget
+import zipfile
+import argparse
+import soundfile
+import json
+from paddle.v2.dataset.common import md5file
+
+DATA_HOME = os.path.expanduser('~/.cache/paddle/dataset/speech')
+
+URL = "https://d4s.myairbridge.com/packagev2/AG0Y3DNBE5IWRRTV/?dlid=W19XG7T0NNHB027139H0EQ"
+MD5 = "c3ff512618d7a67d4f85566ea1bc39ec"
+
+parser = argparse.ArgumentParser(description=__doc__)
+parser.add_argument(
+ "--target_dir",
+ default=DATA_HOME + "/chime3_background",
+ type=str,
+ help="Directory to save the dataset. (default: %(default)s)")
+parser.add_argument(
+ "--manifest_filepath",
+ default="manifest.chime3.background",
+ type=str,
+ help="Filepath for output manifests. (default: %(default)s)")
+args = parser.parse_args()
+
+
+def download(url, md5sum, target_dir, filename=None):
+ """Download file from url to target_dir, and check md5sum."""
+ if filename == None:
+ filename = url.split("/")[-1]
+ if not os.path.exists(target_dir): os.makedirs(target_dir)
+ filepath = os.path.join(target_dir, filename)
+ if not (os.path.exists(filepath) and md5file(filepath) == md5sum):
+ print("Downloading %s ..." % url)
+ wget.download(url, target_dir)
+ print("\nMD5 Chesksum %s ..." % filepath)
+ if not md5file(filepath) == md5sum:
+ raise RuntimeError("MD5 checksum failed.")
+ else:
+ print("File exists, skip downloading. (%s)" % filepath)
+ return filepath
+
+
+def unpack(filepath, target_dir):
+ """Unpack the file to the target_dir."""
+ print("Unpacking %s ..." % filepath)
+ if filepath.endswith('.zip'):
+ zip = zipfile.ZipFile(filepath, 'r')
+ zip.extractall(target_dir)
+ zip.close()
+ elif filepath.endswith('.tar') or filepath.endswith('.tar.gz'):
+ tar = zipfile.open(filepath)
+ tar.extractall(target_dir)
+ tar.close()
+ else:
+ raise ValueError("File format is not supported for unpacking.")
+
+
+def create_manifest(data_dir, manifest_path):
+ """Create a manifest json file summarizing the data set, with each line
+ containing the meta data (i.e. audio filepath, transcription text, audio
+ duration) of each audio file within the data set.
+ """
+ print("Creating manifest %s ..." % manifest_path)
+ json_lines = []
+ for subfolder, _, filelist in sorted(os.walk(data_dir)):
+ for filename in filelist:
+ if filename.endswith('.wav'):
+ filepath = os.path.join(data_dir, subfolder, filename)
+ audio_data, samplerate = soundfile.read(filepath)
+ duration = float(len(audio_data)) / samplerate
+ json_lines.append(
+ json.dumps({
+ 'audio_filepath': filepath,
+ 'duration': duration,
+ 'text': ''
+ }))
+ with open(manifest_path, 'w') as out_file:
+ for line in json_lines:
+ out_file.write(line + '\n')
+
+
+def prepare_chime3(url, md5sum, target_dir, manifest_path):
+ """Download, unpack and create summmary manifest file."""
+ if not os.path.exists(os.path.join(target_dir, "CHiME3")):
+ # download
+ filepath = download(url, md5sum, target_dir,
+ "myairbridge-AG0Y3DNBE5IWRRTV.zip")
+ # unpack
+ unpack(filepath, target_dir)
+ unpack(
+ os.path.join(target_dir, 'CHiME3_background_bus.zip'), target_dir)
+ unpack(
+ os.path.join(target_dir, 'CHiME3_background_caf.zip'), target_dir)
+ unpack(
+ os.path.join(target_dir, 'CHiME3_background_ped.zip'), target_dir)
+ unpack(
+ os.path.join(target_dir, 'CHiME3_background_str.zip'), target_dir)
+ else:
+ print("Skip downloading and unpacking. Data already exists in %s." %
+ target_dir)
+ # create manifest json file
+ create_manifest(target_dir, manifest_path)
+
+
+def main():
+ prepare_chime3(
+ url=URL,
+ md5sum=MD5,
+ target_dir=args.target_dir,
+ manifest_path=args.manifest_filepath)
+
+
+if __name__ == '__main__':
+ main()
diff --git a/data/voxforge/run_data.sh b/data/voxforge/run_data.sh
new file mode 100644
index 000000000..c6ff71118
--- /dev/null
+++ b/data/voxforge/run_data.sh
@@ -0,0 +1,16 @@
+#! /usr/bin/env bash
+
+# download data, generate manifests
+PYTHONPATH=../../:$PYTHONPATH python voxforge.py \
+--manifest_prefix='./manifest' \
+--target_dir='~/.cache/paddle/dataset/speech/VoxForge' \
+--is_merge_dialect=True \
+--dialects 'american' 'british' 'australian' 'european' 'irish' 'canadian' 'indian'
+
+if [ $? -ne 0 ]; then
+ echo "Prepare VoxForge failed. Terminated."
+ exit 1
+fi
+
+echo "VoxForge Data preparation done."
+exit 0
diff --git a/data/voxforge/voxforge.py b/data/voxforge/voxforge.py
new file mode 100644
index 000000000..63f052bd7
--- /dev/null
+++ b/data/voxforge/voxforge.py
@@ -0,0 +1,221 @@
+"""Prepare VoxForge dataset
+
+Download, unpack and create manifest files.
+Manifest file is a json-format file with each line containing the
+meta data (i.e. audio filepath, transcript and audio duration)
+of each audio file in the data set.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import codecs
+import soundfile
+import json
+import argparse
+import shutil
+import subprocess
+from data_utils.utility import download_multi, unpack, getfile_insensitive
+
+DATA_HOME = '~/.cache/paddle/dataset/speech'
+
+DATA_URL = 'http://www.repository.voxforge1.org/downloads/SpeechCorpus/Trunk/' \
+ 'Audio/Main/16kHz_16bit'
+
+parser = argparse.ArgumentParser(description=__doc__)
+parser.add_argument(
+ "--target_dir",
+ default=DATA_HOME + "/VoxForge",
+ type=str,
+ help="Directory to save the dataset. (default: %(default)s)")
+parser.add_argument(
+ "--dialects",
+ default=[
+ 'american', 'british', 'australian', 'european', 'irish', 'canadian',
+ 'indian'
+ ],
+ nargs='+',
+ type=str,
+ help="Dialect types. (default: %(default)s)")
+parser.add_argument(
+ "--is_merge_dialect",
+ default=True,
+ type=bool,
+ help="If set True, manifests of american dialect and canadian dialect will "
+ "be merged to american-canadian dialect; manifests of british "
+ "dialect, irish dialect and australian dialect will be merged to "
+ "commonwealth dialect. (default: %(default)s)")
+parser.add_argument(
+ "--manifest_prefix",
+ default="manifest",
+ type=str,
+ help="Filepath prefix for output manifests. (default: %(default)s)")
+args = parser.parse_args()
+
+
+def download_and_unpack(target_dir, url):
+ wget_args = '-q -l 1 -N -nd -c -e robots=off -A tgz -r -np'
+ tgz_dir = os.path.join(target_dir, 'tgz')
+ exit_code = download_multi(url, tgz_dir, wget_args)
+ if exit_code != 0:
+ print('Download tgz audio files failed with exit code %d.' % exit_code)
+ else:
+ print('Download done, start unpacking ...')
+ audio_dir = os.path.join(target_dir, 'audio')
+ for root, dirs, files in os.walk(tgz_dir):
+ for file in files:
+ print(file)
+ if file.endswith('.tgz'):
+ unpack(os.path.join(root, file), audio_dir)
+
+
+def select_dialects(target_dir, dialect_list):
+ """Classify audio files by dialect."""
+ dialect_root_dir = os.path.join(target_dir, 'dialect')
+ if os.path.exists(dialect_root_dir):
+ shutil.rmtree(dialect_root_dir)
+ os.mkdir(dialect_root_dir)
+ audio_dir = os.path.abspath(os.path.join(target_dir, 'audio'))
+ for dialect in dialect_list:
+ # filter files by dialect
+ command = 'find %s -iwholename "*etc/readme*" -exec egrep -iHl \
+ "pronunciation dialect.*%s" {} \;' % (audio_dir, dialect)
+ p = subprocess.Popen(
+ command, stdin=subprocess.PIPE, stdout=subprocess.PIPE, shell=True)
+ output, err = p.communicate()
+ dialect_dir = os.path.join(dialect_root_dir, dialect)
+ if os.path.exists(dialect_dir):
+ shutil.rmtree(dialect_dir)
+ os.mkdir(dialect_dir)
+ for path in output.splitlines():
+ src_dir = os.path.dirname(os.path.dirname(path))
+ link = os.path.basename(os.path.normpath(src_dir))
+ os.symlink(src_dir, os.path.join(dialect_dir, link))
+
+
+def generate_manifest(data_dir, manifest_path):
+ json_lines = []
+
+ for path in os.listdir(data_dir):
+ audio_link = os.path.join(data_dir, path)
+ assert os.path.islink(
+ audio_link), '%s should be symbolic link.' % audio_link
+ actual_audio_dir = os.path.abspath(os.readlink(audio_link))
+
+ audio_type = ''
+ if os.path.isdir(os.path.join(actual_audio_dir, 'wav')):
+ audio_type = 'wav'
+ elif os.path.isdir(os.path.join(actual_audio_dir, 'flac')):
+ audio_type = 'flac'
+ else:
+ print('Unknown audio type, skipped processing %s.' %
+ actual_audio_dir)
+ continue
+
+ etc_dir = os.path.join(actual_audio_dir, 'etc')
+ prompts_file = os.path.join(etc_dir, 'PROMPTS')
+ if not os.path.isfile(prompts_file):
+ print('PROMPTS file missing, skip processing %s.' %
+ actual_audio_dir)
+ continue
+
+ readme_file = getfile_insensitive(os.path.join(etc_dir, 'README'))
+ if readme_file is None:
+ print('README file missing, skip processing %s.' % actual_audio_dir)
+ continue
+
+ for line in file(prompts_file):
+ u, trans = line.strip().split(None, 1)
+ u_parts = u.split('/')
+
+ # try to format the date time
+ try:
+ speaker, date, sfx = u_parts[-3].split('-')
+ obj = datetime.datetime.strptime(date, '%y.%m.%d')
+ formatted = obj.strftime('%Y%m%d')
+ u_parts[-3] = '-'.join([speaker, formatted, sfx])
+ except Exception as e:
+ pass
+
+ if len(u_parts) < 2:
+ u_parts = [audio_type] + u_parts
+ u_parts[-2] = audio_type
+ u_parts[-1] += '.' + audio_type
+ u = os.path.join(actual_audio_dir, '/'.join(u_parts[-2:]))
+
+ if not os.path.isfile(u):
+ print('Audio file missing, skip processing %s.' % u)
+ continue
+
+ if os.stat(u).st_size == 0:
+ print('Empty audio file, skip processing %s.' % u)
+ continue
+
+ trans = trans.strip().replace('-', ' ')
+ if not trans.isupper() or \
+ not trans.strip().replace(' ', '').replace("'", "").isalpha():
+ print("Transcript not normalized properly, skip processing %s."
+ % u)
+ continue
+
+ audio_data, samplerate = soundfile.read(u)
+ duration = float(len(audio_data)) / samplerate
+ json_lines.append(
+ json.dumps({
+ 'audio_filepath': u,
+ 'duration': duration,
+ 'text': trans.lower()
+ }))
+
+ with codecs.open(manifest_path, 'w', 'utf-8') as fout:
+ for line in json_lines:
+ fout.write(line + '\n')
+
+
+def merge_manifests(manifest_files, save_path):
+ lines = []
+ for manifest_file in manifest_files:
+ line = codecs.open(manifest_file, 'r', 'utf-8').readlines()
+ lines += line
+
+ with codecs.open(save_path, 'w', 'utf-8') as fout:
+ for line in lines:
+ fout.write(line)
+
+
+def prepare_dataset(url, dialects, target_dir, manifest_prefix, is_merge):
+ download_and_unpack(target_dir, url)
+ select_dialects(target_dir, dialects)
+ american_canadian_manifests = []
+ commonwealth_manifests = []
+ for dialect in dialects:
+ dialect_dir = os.path.join(target_dir, 'dialect', dialect)
+ manifest_fpath = manifest_prefix + '.' + dialect
+ if dialect == 'american' or dialect == 'canadian':
+ american_canadian_manifests.append(manifest_fpath)
+ if dialect == 'australian' \
+ or dialect == 'british' \
+ or dialect == 'irish':
+ commonwealth_manifests.append(manifest_fpath)
+ generate_manifest(dialect_dir, manifest_fpath)
+
+ if is_merge:
+ if len(american_canadian_manifests) > 0:
+ manifest_fpath = manifest_prefix + '.american-canadian'
+ merge_manifests(american_canadian_manifests, manifest_fpath)
+ if len(commonwealth_manifests) > 0:
+ manifest_fpath = manifest_prefix + '.commonwealth'
+ merge_manifests(commonwealth_manifests, manifest_fpath)
+
+
+def main():
+ if args.target_dir.startswith('~'):
+ args.target_dir = os.path.expanduser(args.target_dir)
+
+ prepare_dataset(DATA_URL, args.dialects, args.target_dir,
+ args.manifest_prefix, args.is_merge_dialect)
+
+
+if __name__ == '__main__':
+ main()
diff --git a/data_utils/__init__.py b/data_utils/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/data_utils/audio.py b/data_utils/audio.py
new file mode 100644
index 000000000..3fb782951
--- /dev/null
+++ b/data_utils/audio.py
@@ -0,0 +1,685 @@
+"""Contains the audio segment class."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import io
+import struct
+import re
+import soundfile
+import resampy
+from scipy import signal
+import random
+import copy
+
+
+class AudioSegment(object):
+ """Monaural audio segment abstraction.
+
+ :param samples: Audio samples [num_samples x num_channels].
+ :type samples: ndarray.float32
+ :param sample_rate: Audio sample rate.
+ :type sample_rate: int
+ :raises TypeError: If the sample data type is not float or int.
+ """
+
+ def __init__(self, samples, sample_rate):
+ """Create audio segment from samples.
+
+ Samples are convert float32 internally, with int scaled to [-1, 1].
+ """
+ self._samples = self._convert_samples_to_float32(samples)
+ self._sample_rate = sample_rate
+ if self._samples.ndim >= 2:
+ self._samples = np.mean(self._samples, 1)
+
+ def __eq__(self, other):
+ """Return whether two objects are equal."""
+ if type(other) is not type(self):
+ return False
+ if self._sample_rate != other._sample_rate:
+ return False
+ if self._samples.shape != other._samples.shape:
+ return False
+ if np.any(self.samples != other._samples):
+ return False
+ return True
+
+ def __ne__(self, other):
+ """Return whether two objects are unequal."""
+ return not self.__eq__(other)
+
+ def __str__(self):
+ """Return human-readable representation of segment."""
+ return ("%s: num_samples=%d, sample_rate=%d, duration=%.2fsec, "
+ "rms=%.2fdB" % (type(self), self.num_samples, self.sample_rate,
+ self.duration, self.rms_db))
+
+ @classmethod
+ def from_file(cls, file):
+ """Create audio segment from audio file.
+
+ :param filepath: Filepath or file object to audio file.
+ :type filepath: basestring|file
+ :return: Audio segment instance.
+ :rtype: AudioSegment
+ """
+ if isinstance(file, basestring) and re.findall(r".seqbin_\d+$", file):
+ return cls.from_sequence_file(file)
+ else:
+ samples, sample_rate = soundfile.read(file, dtype='float32')
+ return cls(samples, sample_rate)
+
+ @classmethod
+ def slice_from_file(cls, file, start=None, end=None):
+ """Loads a small section of an audio without having to load
+ the entire file into the memory which can be incredibly wasteful.
+
+ :param file: Input audio filepath or file object.
+ :type file: basestring|file
+ :param start: Start time in seconds. If start is negative, it wraps
+ around from the end. If not provided, this function
+ reads from the very beginning.
+ :type start: float
+ :param end: End time in seconds. If end is negative, it wraps around
+ from the end. If not provided, the default behvaior is
+ to read to the end of the file.
+ :type end: float
+ :return: AudioSegment instance of the specified slice of the input
+ audio file.
+ :rtype: AudioSegment
+ :raise ValueError: If start or end is incorrectly set, e.g. out of
+ bounds in time.
+ """
+ sndfile = soundfile.SoundFile(file)
+ sample_rate = sndfile.samplerate
+ duration = float(len(sndfile)) / sample_rate
+ start = 0. if start is None else start
+ end = 0. if end is None else end
+ if start < 0.0:
+ start += duration
+ if end < 0.0:
+ end += duration
+ if start < 0.0:
+ raise ValueError("The slice start position (%f s) is out of "
+ "bounds." % start)
+ if end < 0.0:
+ raise ValueError("The slice end position (%f s) is out of bounds." %
+ end)
+ if start > end:
+ raise ValueError("The slice start position (%f s) is later than "
+ "the slice end position (%f s)." % (start, end))
+ if end > duration:
+ raise ValueError("The slice end position (%f s) is out of bounds "
+ "(> %f s)" % (end, duration))
+ start_frame = int(start * sample_rate)
+ end_frame = int(end * sample_rate)
+ sndfile.seek(start_frame)
+ data = sndfile.read(frames=end_frame - start_frame, dtype='float32')
+ return cls(data, sample_rate)
+
+ @classmethod
+ def from_sequence_file(cls, filepath):
+ """Create audio segment from sequence file. Sequence file is a binary
+ file containing a collection of multiple audio files, with several
+ header bytes in the head indicating the offsets of each audio byte data
+ chunk.
+
+ The format is:
+
+ 4 bytes (int, version),
+ 4 bytes (int, num of utterance),
+ 4 bytes (int, bytes per header),
+ [bytes_per_header*(num_utterance+1)] bytes (offsets for each audio),
+ audio_bytes_data_of_1st_utterance,
+ audio_bytes_data_of_2nd_utterance,
+ ......
+
+ Sequence file name must end with ".seqbin". And the filename of the 5th
+ utterance's audio file in sequence file "xxx.seqbin" must be
+ "xxx.seqbin_5", with "5" indicating the utterance index within this
+ sequence file (starting from 1).
+
+ :param filepath: Filepath of sequence file.
+ :type filepath: basestring
+ :return: Audio segment instance.
+ :rtype: AudioSegment
+ """
+ # parse filepath
+ matches = re.match(r"(.+\.seqbin)_(\d+)", filepath)
+ if matches is None:
+ raise IOError("File type of %s is not supported" % filepath)
+ filename = matches.group(1)
+ fileno = int(matches.group(2))
+
+ # read headers
+ f = open(filename, 'rb')
+ version = f.read(4)
+ num_utterances = struct.unpack("i", f.read(4))[0]
+ bytes_per_header = struct.unpack("i", f.read(4))[0]
+ header_bytes = f.read(bytes_per_header * (num_utterances + 1))
+ header = [
+ struct.unpack("i", header_bytes[bytes_per_header * i:
+ bytes_per_header * (i + 1)])[0]
+ for i in range(num_utterances + 1)
+ ]
+
+ # read audio bytes
+ f.seek(header[fileno - 1])
+ audio_bytes = f.read(header[fileno] - header[fileno - 1])
+ f.close()
+
+ # create audio segment
+ try:
+ return cls.from_bytes(audio_bytes)
+ except Exception as e:
+ samples = np.frombuffer(audio_bytes, dtype='int16')
+ return cls(samples=samples, sample_rate=8000)
+
+ @classmethod
+ def from_bytes(cls, bytes):
+ """Create audio segment from a byte string containing audio samples.
+
+ :param bytes: Byte string containing audio samples.
+ :type bytes: str
+ :return: Audio segment instance.
+ :rtype: AudioSegment
+ """
+ samples, sample_rate = soundfile.read(
+ io.BytesIO(bytes), dtype='float32')
+ return cls(samples, sample_rate)
+
+ @classmethod
+ def concatenate(cls, *segments):
+ """Concatenate an arbitrary number of audio segments together.
+
+ :param *segments: Input audio segments to be concatenated.
+ :type *segments: tuple of AudioSegment
+ :return: Audio segment instance as concatenating results.
+ :rtype: AudioSegment
+ :raises ValueError: If the number of segments is zero, or if the
+ sample_rate of any segments does not match.
+ :raises TypeError: If any segment is not AudioSegment instance.
+ """
+ # Perform basic sanity-checks.
+ if len(segments) == 0:
+ raise ValueError("No audio segments are given to concatenate.")
+ sample_rate = segments[0]._sample_rate
+ for seg in segments:
+ if sample_rate != seg._sample_rate:
+ raise ValueError("Can't concatenate segments with "
+ "different sample rates")
+ if type(seg) is not cls:
+ raise TypeError("Only audio segments of the same type "
+ "can be concatenated.")
+ samples = np.concatenate([seg.samples for seg in segments])
+ return cls(samples, sample_rate)
+
+ @classmethod
+ def make_silence(cls, duration, sample_rate):
+ """Creates a silent audio segment of the given duration and sample rate.
+
+ :param duration: Length of silence in seconds.
+ :type duration: float
+ :param sample_rate: Sample rate.
+ :type sample_rate: float
+ :return: Silent AudioSegment instance of the given duration.
+ :rtype: AudioSegment
+ """
+ samples = np.zeros(int(duration * sample_rate))
+ return cls(samples, sample_rate)
+
+ def to_wav_file(self, filepath, dtype='float32'):
+ """Save audio segment to disk as wav file.
+
+ :param filepath: WAV filepath or file object to save the
+ audio segment.
+ :type filepath: basestring|file
+ :param dtype: Subtype for audio file. Options: 'int16', 'int32',
+ 'float32', 'float64'. Default is 'float32'.
+ :type dtype: str
+ :raises TypeError: If dtype is not supported.
+ """
+ samples = self._convert_samples_from_float32(self._samples, dtype)
+ subtype_map = {
+ 'int16': 'PCM_16',
+ 'int32': 'PCM_32',
+ 'float32': 'FLOAT',
+ 'float64': 'DOUBLE'
+ }
+ soundfile.write(
+ filepath,
+ samples,
+ self._sample_rate,
+ format='WAV',
+ subtype=subtype_map[dtype])
+
+ def superimpose(self, other):
+ """Add samples from another segment to those of this segment
+ (sample-wise addition, not segment concatenation).
+
+ Note that this is an in-place transformation.
+
+ :param other: Segment containing samples to be added in.
+ :type other: AudioSegments
+ :raise TypeError: If type of two segments don't match.
+ :raise ValueError: If the sample rates of the two segments are not
+ equal, or if the lengths of segments don't match.
+ """
+ if isinstance(other, type(self)):
+ raise TypeError("Cannot add segments of different types: %s "
+ "and %s." % (type(self), type(other)))
+ if self._sample_rate != other._sample_rate:
+ raise ValueError("Sample rates must match to add segments.")
+ if len(self._samples) != len(other._samples):
+ raise ValueError("Segment lengths must match to add segments.")
+ self._samples += other._samples
+
+ def to_bytes(self, dtype='float32'):
+ """Create a byte string containing the audio content.
+
+ :param dtype: Data type for export samples. Options: 'int16', 'int32',
+ 'float32', 'float64'. Default is 'float32'.
+ :type dtype: str
+ :return: Byte string containing audio content.
+ :rtype: str
+ """
+ samples = self._convert_samples_from_float32(self._samples, dtype)
+ return samples.tostring()
+
+ def gain_db(self, gain):
+ """Apply gain in decibels to samples.
+
+ Note that this is an in-place transformation.
+
+ :param gain: Gain in decibels to apply to samples.
+ :type gain: float|1darray
+ """
+ self._samples *= 10.**(gain / 20.)
+
+ def change_speed(self, speed_rate):
+ """Change the audio speed by linear interpolation.
+
+ Note that this is an in-place transformation.
+
+ :param speed_rate: Rate of speed change:
+ speed_rate > 1.0, speed up the audio;
+ speed_rate = 1.0, unchanged;
+ speed_rate < 1.0, slow down the audio;
+ speed_rate <= 0.0, not allowed, raise ValueError.
+ :type speed_rate: float
+ :raises ValueError: If speed_rate <= 0.0.
+ """
+ if speed_rate <= 0:
+ raise ValueError("speed_rate should be greater than zero.")
+ old_length = self._samples.shape[0]
+ new_length = int(old_length / speed_rate)
+ old_indices = np.arange(old_length)
+ new_indices = np.linspace(start=0, stop=old_length, num=new_length)
+ self._samples = np.interp(new_indices, old_indices, self._samples)
+
+ def normalize(self, target_db=-20, max_gain_db=300.0):
+ """Normalize audio to be of the desired RMS value in decibels.
+
+ Note that this is an in-place transformation.
+
+ :param target_db: Target RMS value in decibels. This value should be
+ less than 0.0 as 0.0 is full-scale audio.
+ :type target_db: float
+ :param max_gain_db: Max amount of gain in dB that can be applied for
+ normalization. This is to prevent nans when
+ attempting to normalize a signal consisting of
+ all zeros.
+ :type max_gain_db: float
+ :raises ValueError: If the required gain to normalize the segment to
+ the target_db value exceeds max_gain_db.
+ """
+ gain = target_db - self.rms_db
+ if gain > max_gain_db:
+ raise ValueError(
+ "Unable to normalize segment to %f dB because the "
+ "the probable gain have exceeds max_gain_db (%f dB)" %
+ (target_db, max_gain_db))
+ self.gain_db(min(max_gain_db, target_db - self.rms_db))
+
+ def normalize_online_bayesian(self,
+ target_db,
+ prior_db,
+ prior_samples,
+ startup_delay=0.0):
+ """Normalize audio using a production-compatible online/causal
+ algorithm. This uses an exponential likelihood and gamma prior to
+ make online estimates of the RMS even when there are very few samples.
+
+ Note that this is an in-place transformation.
+
+ :param target_db: Target RMS value in decibels.
+ :type target_bd: float
+ :param prior_db: Prior RMS estimate in decibels.
+ :type prior_db: float
+ :param prior_samples: Prior strength in number of samples.
+ :type prior_samples: float
+ :param startup_delay: Default 0.0s. If provided, this function will
+ accrue statistics for the first startup_delay
+ seconds before applying online normalization.
+ :type startup_delay: float
+ """
+ # Estimate total RMS online.
+ startup_sample_idx = min(self.num_samples - 1,
+ int(self.sample_rate * startup_delay))
+ prior_mean_squared = 10.**(prior_db / 10.)
+ prior_sum_of_squares = prior_mean_squared * prior_samples
+ cumsum_of_squares = np.cumsum(self.samples**2)
+ sample_count = np.arange(self.num_samples) + 1
+ if startup_sample_idx > 0:
+ cumsum_of_squares[:startup_sample_idx] = \
+ cumsum_of_squares[startup_sample_idx]
+ sample_count[:startup_sample_idx] = \
+ sample_count[startup_sample_idx]
+ mean_squared_estimate = ((cumsum_of_squares + prior_sum_of_squares) /
+ (sample_count + prior_samples))
+ rms_estimate_db = 10 * np.log10(mean_squared_estimate)
+ # Compute required time-varying gain.
+ gain_db = target_db - rms_estimate_db
+ self.gain_db(gain_db)
+
+ def resample(self, target_sample_rate, filter='kaiser_best'):
+ """Resample the audio to a target sample rate.
+
+ Note that this is an in-place transformation.
+
+ :param target_sample_rate: Target sample rate.
+ :type target_sample_rate: int
+ :param filter: The resampling filter to use one of {'kaiser_best',
+ 'kaiser_fast'}.
+ :type filter: str
+ """
+ self._samples = resampy.resample(
+ self.samples, self.sample_rate, target_sample_rate, filter=filter)
+ self._sample_rate = target_sample_rate
+
+ def pad_silence(self, duration, sides='both'):
+ """Pad this audio sample with a period of silence.
+
+ Note that this is an in-place transformation.
+
+ :param duration: Length of silence in seconds to pad.
+ :type duration: float
+ :param sides: Position for padding:
+ 'beginning' - adds silence in the beginning;
+ 'end' - adds silence in the end;
+ 'both' - adds silence in both the beginning and the end.
+ :type sides: str
+ :raises ValueError: If sides is not supported.
+ """
+ if duration == 0.0:
+ return self
+ cls = type(self)
+ silence = self.make_silence(duration, self._sample_rate)
+ if sides == "beginning":
+ padded = cls.concatenate(silence, self)
+ elif sides == "end":
+ padded = cls.concatenate(self, silence)
+ elif sides == "both":
+ padded = cls.concatenate(silence, self, silence)
+ else:
+ raise ValueError("Unknown value for the sides %s" % sides)
+ self._samples = padded._samples
+
+ def shift(self, shift_ms):
+ """Shift the audio in time. If `shift_ms` is positive, shift with time
+ advance; if negative, shift with time delay. Silence are padded to
+ keep the duration unchanged.
+
+ Note that this is an in-place transformation.
+
+ :param shift_ms: Shift time in millseconds. If positive, shift with
+ time advance; if negative; shift with time delay.
+ :type shift_ms: float
+ :raises ValueError: If shift_ms is longer than audio duration.
+ """
+ if abs(shift_ms) / 1000.0 > self.duration:
+ raise ValueError("Absolute value of shift_ms should be smaller "
+ "than audio duration.")
+ shift_samples = int(shift_ms * self._sample_rate / 1000)
+ if shift_samples > 0:
+ # time advance
+ self._samples[:-shift_samples] = self._samples[shift_samples:]
+ self._samples[-shift_samples:] = 0
+ elif shift_samples < 0:
+ # time delay
+ self._samples[-shift_samples:] = self._samples[:shift_samples]
+ self._samples[:-shift_samples] = 0
+
+ def subsegment(self, start_sec=None, end_sec=None):
+ """Cut the AudioSegment between given boundaries.
+
+ Note that this is an in-place transformation.
+
+ :param start_sec: Beginning of subsegment in seconds.
+ :type start_sec: float
+ :param end_sec: End of subsegment in seconds.
+ :type end_sec: float
+ :raise ValueError: If start_sec or end_sec is incorrectly set, e.g. out
+ of bounds in time.
+ """
+ start_sec = 0.0 if start_sec is None else start_sec
+ end_sec = self.duration if end_sec is None else end_sec
+ if start_sec < 0.0:
+ start_sec = self.duration + start_sec
+ if end_sec < 0.0:
+ end_sec = self.duration + end_sec
+ if start_sec < 0.0:
+ raise ValueError("The slice start position (%f s) is out of "
+ "bounds." % start_sec)
+ if end_sec < 0.0:
+ raise ValueError("The slice end position (%f s) is out of bounds." %
+ end_sec)
+ if start_sec > end_sec:
+ raise ValueError("The slice start position (%f s) is later than "
+ "the end position (%f s)." % (start_sec, end_sec))
+ if end_sec > self.duration:
+ raise ValueError("The slice end position (%f s) is out of bounds "
+ "(> %f s)" % (end_sec, self.duration))
+ start_sample = int(round(start_sec * self._sample_rate))
+ end_sample = int(round(end_sec * self._sample_rate))
+ self._samples = self._samples[start_sample:end_sample]
+
+ def random_subsegment(self, subsegment_length, rng=None):
+ """Cut the specified length of the audiosegment randomly.
+
+ Note that this is an in-place transformation.
+
+ :param subsegment_length: Subsegment length in seconds.
+ :type subsegment_length: float
+ :param rng: Random number generator state.
+ :type rng: random.Random
+ :raises ValueError: If the length of subsegment is greater than
+ the origineal segemnt.
+ """
+ rng = random.Random() if rng is None else rng
+ if subsegment_length > self.duration:
+ raise ValueError("Length of subsegment must not be greater "
+ "than original segment.")
+ start_time = rng.uniform(0.0, self.duration - subsegment_length)
+ self.subsegment(start_time, start_time + subsegment_length)
+
+ def convolve(self, impulse_segment, allow_resample=False):
+ """Convolve this audio segment with the given impulse segment.
+
+ Note that this is an in-place transformation.
+
+ :param impulse_segment: Impulse response segments.
+ :type impulse_segment: AudioSegment
+ :param allow_resample: Indicates whether resampling is allowed when
+ the impulse_segment has a different sample
+ rate from this signal.
+ :type allow_resample: bool
+ :raises ValueError: If the sample rate is not match between two
+ audio segments when resample is not allowed.
+ """
+ if allow_resample and self.sample_rate != impulse_segment.sample_rate:
+ impulse_segment.resample(self.sample_rate)
+ if self.sample_rate != impulse_segment.sample_rate:
+ raise ValueError("Impulse segment's sample rate (%d Hz) is not "
+ "equal to base signal sample rate (%d Hz)." %
+ (impulse_segment.sample_rate, self.sample_rate))
+ samples = signal.fftconvolve(self.samples, impulse_segment.samples,
+ "full")
+ self._samples = samples
+
+ def convolve_and_normalize(self, impulse_segment, allow_resample=False):
+ """Convolve and normalize the resulting audio segment so that it
+ has the same average power as the input signal.
+
+ Note that this is an in-place transformation.
+
+ :param impulse_segment: Impulse response segments.
+ :type impulse_segment: AudioSegment
+ :param allow_resample: Indicates whether resampling is allowed when
+ the impulse_segment has a different sample
+ rate from this signal.
+ :type allow_resample: bool
+ """
+ target_db = self.rms_db
+ self.convolve(impulse_segment, allow_resample=allow_resample)
+ self.normalize(target_db)
+
+ def add_noise(self,
+ noise,
+ snr_dB,
+ allow_downsampling=False,
+ max_gain_db=300.0,
+ rng=None):
+ """Add the given noise segment at a specific signal-to-noise ratio.
+ If the noise segment is longer than this segment, a random subsegment
+ of matching length is sampled from it and used instead.
+
+ Note that this is an in-place transformation.
+
+ :param noise: Noise signal to add.
+ :type noise: AudioSegment
+ :param snr_dB: Signal-to-Noise Ratio, in decibels.
+ :type snr_dB: float
+ :param allow_downsampling: Whether to allow the noise signal to be
+ downsampled to match the base signal sample
+ rate.
+ :type allow_downsampling: bool
+ :param max_gain_db: Maximum amount of gain to apply to noise signal
+ before adding it in. This is to prevent attempting
+ to apply infinite gain to a zero signal.
+ :type max_gain_db: float
+ :param rng: Random number generator state.
+ :type rng: None|random.Random
+ :raises ValueError: If the sample rate does not match between the two
+ audio segments when downsampling is not allowed, or
+ if the duration of noise segments is shorter than
+ original audio segments.
+ """
+ rng = random.Random() if rng is None else rng
+ if allow_downsampling and noise.sample_rate > self.sample_rate:
+ noise = noise.resample(self.sample_rate)
+ if noise.sample_rate != self.sample_rate:
+ raise ValueError("Noise sample rate (%d Hz) is not equal to base "
+ "signal sample rate (%d Hz)." % (noise.sample_rate,
+ self.sample_rate))
+ if noise.duration < self.duration:
+ raise ValueError("Noise signal (%f sec) must be at least as long as"
+ " base signal (%f sec)." %
+ (noise.duration, self.duration))
+ noise_gain_db = min(self.rms_db - noise.rms_db - snr_dB, max_gain_db)
+ noise_new = copy.deepcopy(noise)
+ noise_new.random_subsegment(self.duration, rng=rng)
+ noise_new.gain_db(noise_gain_db)
+ self.superimpose(noise_new)
+
+ @property
+ def samples(self):
+ """Return audio samples.
+
+ :return: Audio samples.
+ :rtype: ndarray
+ """
+ return self._samples.copy()
+
+ @property
+ def sample_rate(self):
+ """Return audio sample rate.
+
+ :return: Audio sample rate.
+ :rtype: int
+ """
+ return self._sample_rate
+
+ @property
+ def num_samples(self):
+ """Return number of samples.
+
+ :return: Number of samples.
+ :rtype: int
+ """
+ return self._samples.shape[0]
+
+ @property
+ def duration(self):
+ """Return audio duration.
+
+ :return: Audio duration in seconds.
+ :rtype: float
+ """
+ return self._samples.shape[0] / float(self._sample_rate)
+
+ @property
+ def rms_db(self):
+ """Return root mean square energy of the audio in decibels.
+
+ :return: Root mean square energy in decibels.
+ :rtype: float
+ """
+ # square root => multiply by 10 instead of 20 for dBs
+ mean_square = np.mean(self._samples**2)
+ return 10 * np.log10(mean_square)
+
+ def _convert_samples_to_float32(self, samples):
+ """Convert sample type to float32.
+
+ Audio sample type is usually integer or float-point.
+ Integers will be scaled to [-1, 1] in float32.
+ """
+ float32_samples = samples.astype('float32')
+ if samples.dtype in np.sctypes['int']:
+ bits = np.iinfo(samples.dtype).bits
+ float32_samples *= (1. / 2**(bits - 1))
+ elif samples.dtype in np.sctypes['float']:
+ pass
+ else:
+ raise TypeError("Unsupported sample type: %s." % samples.dtype)
+ return float32_samples
+
+ def _convert_samples_from_float32(self, samples, dtype):
+ """Convert sample type from float32 to dtype.
+
+ Audio sample type is usually integer or float-point. For integer
+ type, float32 will be rescaled from [-1, 1] to the maximum range
+ supported by the integer type.
+
+ This is for writing a audio file.
+ """
+ dtype = np.dtype(dtype)
+ output_samples = samples.copy()
+ if dtype in np.sctypes['int']:
+ bits = np.iinfo(dtype).bits
+ output_samples *= (2**(bits - 1) / 1.)
+ min_val = np.iinfo(dtype).min
+ max_val = np.iinfo(dtype).max
+ output_samples[output_samples > max_val] = max_val
+ output_samples[output_samples < min_val] = min_val
+ elif samples.dtype in np.sctypes['float']:
+ min_val = np.finfo(dtype).min
+ max_val = np.finfo(dtype).max
+ output_samples[output_samples > max_val] = max_val
+ output_samples[output_samples < min_val] = min_val
+ else:
+ raise TypeError("Unsupported sample type: %s." % samples.dtype)
+ return output_samples.astype(dtype)
diff --git a/data_utils/augmentor/__init__.py b/data_utils/augmentor/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/data_utils/augmentor/augmentation.py b/data_utils/augmentor/augmentation.py
new file mode 100644
index 000000000..5c30b627e
--- /dev/null
+++ b/data_utils/augmentor/augmentation.py
@@ -0,0 +1,124 @@
+"""Contains the data augmentation pipeline."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import json
+import random
+from data_utils.augmentor.volume_perturb import VolumePerturbAugmentor
+from data_utils.augmentor.shift_perturb import ShiftPerturbAugmentor
+from data_utils.augmentor.speed_perturb import SpeedPerturbAugmentor
+from data_utils.augmentor.noise_perturb import NoisePerturbAugmentor
+from data_utils.augmentor.impulse_response import ImpulseResponseAugmentor
+from data_utils.augmentor.resample import ResampleAugmentor
+from data_utils.augmentor.online_bayesian_normalization import \
+ OnlineBayesianNormalizationAugmentor
+
+
+class AugmentationPipeline(object):
+ """Build a pre-processing pipeline with various augmentation models.Such a
+ data augmentation pipeline is oftern leveraged to augment the training
+ samples to make the model invariant to certain types of perturbations in the
+ real world, improving model's generalization ability.
+
+ The pipeline is built according the the augmentation configuration in json
+ string, e.g.
+
+ .. code-block::
+
+ [ {
+ "type": "noise",
+ "params": {"min_snr_dB": 10,
+ "max_snr_dB": 20,
+ "noise_manifest_path": "datasets/manifest.noise"},
+ "prob": 0.0
+ },
+ {
+ "type": "speed",
+ "params": {"min_speed_rate": 0.9,
+ "max_speed_rate": 1.1},
+ "prob": 1.0
+ },
+ {
+ "type": "shift",
+ "params": {"min_shift_ms": -5,
+ "max_shift_ms": 5},
+ "prob": 1.0
+ },
+ {
+ "type": "volume",
+ "params": {"min_gain_dBFS": -10,
+ "max_gain_dBFS": 10},
+ "prob": 0.0
+ },
+ {
+ "type": "bayesian_normal",
+ "params": {"target_db": -20,
+ "prior_db": -20,
+ "prior_samples": 100},
+ "prob": 0.0
+ }
+ ]
+
+ This augmentation configuration inserts two augmentation models
+ into the pipeline, with one is VolumePerturbAugmentor and the other
+ SpeedPerturbAugmentor. "prob" indicates the probability of the current
+ augmentor to take effect. If "prob" is zero, the augmentor does not take
+ effect.
+
+ :param augmentation_config: Augmentation configuration in json string.
+ :type augmentation_config: str
+ :param random_seed: Random seed.
+ :type random_seed: int
+ :raises ValueError: If the augmentation json config is in incorrect format".
+ """
+
+ def __init__(self, augmentation_config, random_seed=0):
+ self._rng = random.Random(random_seed)
+ self._augmentors, self._rates = self._parse_pipeline_from(
+ augmentation_config)
+
+ def transform_audio(self, audio_segment):
+ """Run the pre-processing pipeline for data augmentation.
+
+ Note that this is an in-place transformation.
+
+ :param audio_segment: Audio segment to process.
+ :type audio_segment: AudioSegmenet|SpeechSegment
+ """
+ for augmentor, rate in zip(self._augmentors, self._rates):
+ if self._rng.uniform(0., 1.) < rate:
+ augmentor.transform_audio(audio_segment)
+
+ def _parse_pipeline_from(self, config_json):
+ """Parse the config json to build a augmentation pipelien."""
+ try:
+ configs = json.loads(config_json)
+ augmentors = [
+ self._get_augmentor(config["type"], config["params"])
+ for config in configs
+ ]
+ rates = [config["prob"] for config in configs]
+ except Exception as e:
+ raise ValueError("Failed to parse the augmentation config json: "
+ "%s" % str(e))
+ return augmentors, rates
+
+ def _get_augmentor(self, augmentor_type, params):
+ """Return an augmentation model by the type name, and pass in params."""
+ if augmentor_type == "volume":
+ return VolumePerturbAugmentor(self._rng, **params)
+ elif augmentor_type == "shift":
+ return ShiftPerturbAugmentor(self._rng, **params)
+ elif augmentor_type == "speed":
+ return SpeedPerturbAugmentor(self._rng, **params)
+ elif augmentor_type == "resample":
+ return ResampleAugmentor(self._rng, **params)
+ elif augmentor_type == "bayesian_normal":
+ return OnlineBayesianNormalizationAugmentor(self._rng, **params)
+ elif augmentor_type == "noise":
+ return NoisePerturbAugmentor(self._rng, **params)
+ elif augmentor_type == "impulse":
+ return ImpulseResponseAugmentor(self._rng, **params)
+ else:
+ raise ValueError("Unknown augmentor type [%s]." % augmentor_type)
diff --git a/data_utils/augmentor/base.py b/data_utils/augmentor/base.py
new file mode 100644
index 000000000..a323165aa
--- /dev/null
+++ b/data_utils/augmentor/base.py
@@ -0,0 +1,33 @@
+"""Contains the abstract base class for augmentation models."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from abc import ABCMeta, abstractmethod
+
+
+class AugmentorBase(object):
+ """Abstract base class for augmentation model (augmentor) class.
+ All augmentor classes should inherit from this class, and implement the
+ following abstract methods.
+ """
+
+ __metaclass__ = ABCMeta
+
+ @abstractmethod
+ def __init__(self):
+ pass
+
+ @abstractmethod
+ def transform_audio(self, audio_segment):
+ """Adds various effects to the input audio segment. Such effects
+ will augment the training data to make the model invariant to certain
+ types of perturbations in the real world, improving model's
+ generalization ability.
+
+ Note that this is an in-place transformation.
+
+ :param audio_segment: Audio segment to add effects to.
+ :type audio_segment: AudioSegmenet|SpeechSegment
+ """
+ pass
diff --git a/data_utils/augmentor/impulse_response.py b/data_utils/augmentor/impulse_response.py
new file mode 100644
index 000000000..536b4d6a4
--- /dev/null
+++ b/data_utils/augmentor/impulse_response.py
@@ -0,0 +1,34 @@
+"""Contains the impulse response augmentation model."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from data_utils.augmentor.base import AugmentorBase
+from data_utils.utility import read_manifest
+from data_utils.audio import AudioSegment
+
+
+class ImpulseResponseAugmentor(AugmentorBase):
+ """Augmentation model for adding impulse response effect.
+
+ :param rng: Random generator object.
+ :type rng: random.Random
+ :param impulse_manifest_path: Manifest path for impulse audio data.
+ :type impulse_manifest_path: basestring
+ """
+
+ def __init__(self, rng, impulse_manifest_path):
+ self._rng = rng
+ self._impulse_manifest = read_manifest(impulse_manifest_path)
+
+ def transform_audio(self, audio_segment):
+ """Add impulse response effect.
+
+ Note that this is an in-place transformation.
+
+ :param audio_segment: Audio segment to add effects to.
+ :type audio_segment: AudioSegmenet|SpeechSegment
+ """
+ impulse_json = self._rng.sample(self._impulse_manifest, 1)[0]
+ impulse_segment = AudioSegment.from_file(impulse_json['audio_filepath'])
+ audio_segment.convolve(impulse_segment, allow_resample=True)
diff --git a/data_utils/augmentor/noise_perturb.py b/data_utils/augmentor/noise_perturb.py
new file mode 100644
index 000000000..96e0ff4de
--- /dev/null
+++ b/data_utils/augmentor/noise_perturb.py
@@ -0,0 +1,49 @@
+"""Contains the noise perturb augmentation model."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from data_utils.augmentor.base import AugmentorBase
+from data_utils.utility import read_manifest
+from data_utils.audio import AudioSegment
+
+
+class NoisePerturbAugmentor(AugmentorBase):
+ """Augmentation model for adding background noise.
+
+ :param rng: Random generator object.
+ :type rng: random.Random
+ :param min_snr_dB: Minimal signal noise ratio, in decibels.
+ :type min_snr_dB: float
+ :param max_snr_dB: Maximal signal noise ratio, in decibels.
+ :type max_snr_dB: float
+ :param noise_manifest_path: Manifest path for noise audio data.
+ :type noise_manifest_path: basestring
+ """
+
+ def __init__(self, rng, min_snr_dB, max_snr_dB, noise_manifest_path):
+ self._min_snr_dB = min_snr_dB
+ self._max_snr_dB = max_snr_dB
+ self._rng = rng
+ self._noise_manifest = read_manifest(manifest_path=noise_manifest_path)
+
+ def transform_audio(self, audio_segment):
+ """Add background noise audio.
+
+ Note that this is an in-place transformation.
+
+ :param audio_segment: Audio segment to add effects to.
+ :type audio_segment: AudioSegmenet|SpeechSegment
+ """
+ noise_json = self._rng.sample(self._noise_manifest, 1)[0]
+ if noise_json['duration'] < audio_segment.duration:
+ raise RuntimeError("The duration of sampled noise audio is smaller "
+ "than the audio segment to add effects to.")
+ diff_duration = noise_json['duration'] - audio_segment.duration
+ start = self._rng.uniform(0, diff_duration)
+ end = start + audio_segment.duration
+ noise_segment = AudioSegment.slice_from_file(
+ noise_json['audio_filepath'], start=start, end=end)
+ snr_dB = self._rng.uniform(self._min_snr_dB, self._max_snr_dB)
+ audio_segment.add_noise(
+ noise_segment, snr_dB, allow_downsampling=True, rng=self._rng)
diff --git a/data_utils/augmentor/online_bayesian_normalization.py b/data_utils/augmentor/online_bayesian_normalization.py
new file mode 100644
index 000000000..e488ac7d6
--- /dev/null
+++ b/data_utils/augmentor/online_bayesian_normalization.py
@@ -0,0 +1,48 @@
+"""Contain the online bayesian normalization augmentation model."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from data_utils.augmentor.base import AugmentorBase
+
+
+class OnlineBayesianNormalizationAugmentor(AugmentorBase):
+ """Augmentation model for adding online bayesian normalization.
+
+ :param rng: Random generator object.
+ :type rng: random.Random
+ :param target_db: Target RMS value in decibels.
+ :type target_db: float
+ :param prior_db: Prior RMS estimate in decibels.
+ :type prior_db: float
+ :param prior_samples: Prior strength in number of samples.
+ :type prior_samples: int
+ :param startup_delay: Default 0.0s. If provided, this function will
+ accrue statistics for the first startup_delay
+ seconds before applying online normalization.
+ :type starup_delay: float.
+ """
+
+ def __init__(self,
+ rng,
+ target_db,
+ prior_db,
+ prior_samples,
+ startup_delay=0.0):
+ self._target_db = target_db
+ self._prior_db = prior_db
+ self._prior_samples = prior_samples
+ self._rng = rng
+ self._startup_delay = startup_delay
+
+ def transform_audio(self, audio_segment):
+ """Normalizes the input audio using the online Bayesian approach.
+
+ Note that this is an in-place transformation.
+
+ :param audio_segment: Audio segment to add effects to.
+ :type audio_segment: AudioSegment|SpeechSegment
+ """
+ audio_segment.normalize_online_bayesian(self._target_db, self._prior_db,
+ self._prior_samples,
+ self._startup_delay)
diff --git a/data_utils/augmentor/resample.py b/data_utils/augmentor/resample.py
new file mode 100644
index 000000000..8df17f3a8
--- /dev/null
+++ b/data_utils/augmentor/resample.py
@@ -0,0 +1,33 @@
+"""Contain the resample augmentation model."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from data_utils.augmentor.base import AugmentorBase
+
+
+class ResampleAugmentor(AugmentorBase):
+ """Augmentation model for resampling.
+
+ See more info here:
+ https://ccrma.stanford.edu/~jos/resample/index.html
+
+ :param rng: Random generator object.
+ :type rng: random.Random
+ :param new_sample_rate: New sample rate in Hz.
+ :type new_sample_rate: int
+ """
+
+ def __init__(self, rng, new_sample_rate):
+ self._new_sample_rate = new_sample_rate
+ self._rng = rng
+
+ def transform_audio(self, audio_segment):
+ """Resamples the input audio to a target sample rate.
+
+ Note that this is an in-place transformation.
+
+ :param audio: Audio segment to add effects to.
+ :type audio: AudioSegment|SpeechSegment
+ """
+ audio_segment.resample(self._new_sample_rate)
diff --git a/data_utils/augmentor/shift_perturb.py b/data_utils/augmentor/shift_perturb.py
new file mode 100644
index 000000000..c4cbe3e17
--- /dev/null
+++ b/data_utils/augmentor/shift_perturb.py
@@ -0,0 +1,34 @@
+"""Contains the volume perturb augmentation model."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from data_utils.augmentor.base import AugmentorBase
+
+
+class ShiftPerturbAugmentor(AugmentorBase):
+ """Augmentation model for adding random shift perturbation.
+
+ :param rng: Random generator object.
+ :type rng: random.Random
+ :param min_shift_ms: Minimal shift in milliseconds.
+ :type min_shift_ms: float
+ :param max_shift_ms: Maximal shift in milliseconds.
+ :type max_shift_ms: float
+ """
+
+ def __init__(self, rng, min_shift_ms, max_shift_ms):
+ self._min_shift_ms = min_shift_ms
+ self._max_shift_ms = max_shift_ms
+ self._rng = rng
+
+ def transform_audio(self, audio_segment):
+ """Shift audio.
+
+ Note that this is an in-place transformation.
+
+ :param audio_segment: Audio segment to add effects to.
+ :type audio_segment: AudioSegmenet|SpeechSegment
+ """
+ shift_ms = self._rng.uniform(self._min_shift_ms, self._max_shift_ms)
+ audio_segment.shift(shift_ms)
diff --git a/data_utils/augmentor/speed_perturb.py b/data_utils/augmentor/speed_perturb.py
new file mode 100644
index 000000000..cc5738bd1
--- /dev/null
+++ b/data_utils/augmentor/speed_perturb.py
@@ -0,0 +1,47 @@
+"""Contain the speech perturbation augmentation model."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from data_utils.augmentor.base import AugmentorBase
+
+
+class SpeedPerturbAugmentor(AugmentorBase):
+ """Augmentation model for adding speed perturbation.
+
+ See reference paper here:
+ http://www.danielpovey.com/files/2015_interspeech_augmentation.pdf
+
+ :param rng: Random generator object.
+ :type rng: random.Random
+ :param min_speed_rate: Lower bound of new speed rate to sample and should
+ not be smaller than 0.9.
+ :type min_speed_rate: float
+ :param max_speed_rate: Upper bound of new speed rate to sample and should
+ not be larger than 1.1.
+ :type max_speed_rate: float
+ """
+
+ def __init__(self, rng, min_speed_rate, max_speed_rate):
+ if min_speed_rate < 0.9:
+ raise ValueError(
+ "Sampling speed below 0.9 can cause unnatural effects")
+ if max_speed_rate > 1.1:
+ raise ValueError(
+ "Sampling speed above 1.1 can cause unnatural effects")
+ self._min_speed_rate = min_speed_rate
+ self._max_speed_rate = max_speed_rate
+ self._rng = rng
+
+ def transform_audio(self, audio_segment):
+ """Sample a new speed rate from the given range and
+ changes the speed of the given audio clip.
+
+ Note that this is an in-place transformation.
+
+ :param audio_segment: Audio segment to add effects to.
+ :type audio_segment: AudioSegment|SpeechSegment
+ """
+ sampled_speed = self._rng.uniform(self._min_speed_rate,
+ self._max_speed_rate)
+ audio_segment.change_speed(sampled_speed)
diff --git a/data_utils/augmentor/volume_perturb.py b/data_utils/augmentor/volume_perturb.py
new file mode 100644
index 000000000..758676d55
--- /dev/null
+++ b/data_utils/augmentor/volume_perturb.py
@@ -0,0 +1,40 @@
+"""Contains the volume perturb augmentation model."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from data_utils.augmentor.base import AugmentorBase
+
+
+class VolumePerturbAugmentor(AugmentorBase):
+ """Augmentation model for adding random volume perturbation.
+
+ This is used for multi-loudness training of PCEN. See
+
+ https://arxiv.org/pdf/1607.05666v1.pdf
+
+ for more details.
+
+ :param rng: Random generator object.
+ :type rng: random.Random
+ :param min_gain_dBFS: Minimal gain in dBFS.
+ :type min_gain_dBFS: float
+ :param max_gain_dBFS: Maximal gain in dBFS.
+ :type max_gain_dBFS: float
+ """
+
+ def __init__(self, rng, min_gain_dBFS, max_gain_dBFS):
+ self._min_gain_dBFS = min_gain_dBFS
+ self._max_gain_dBFS = max_gain_dBFS
+ self._rng = rng
+
+ def transform_audio(self, audio_segment):
+ """Change audio loadness.
+
+ Note that this is an in-place transformation.
+
+ :param audio_segment: Audio segment to add effects to.
+ :type audio_segment: AudioSegmenet|SpeechSegment
+ """
+ gain = self._rng.uniform(self._min_gain_dBFS, self._max_gain_dBFS)
+ audio_segment.gain_db(gain)
diff --git a/data_utils/data.py b/data_utils/data.py
new file mode 100644
index 000000000..f79a6395a
--- /dev/null
+++ b/data_utils/data.py
@@ -0,0 +1,346 @@
+"""Contains data generator for orgnaizing various audio data preprocessing
+pipeline and offering data reader interface of PaddlePaddle requirements.
+"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import random
+import tarfile
+import multiprocessing
+import numpy as np
+import paddle.v2 as paddle
+from threading import local
+from data_utils.utility import read_manifest
+from data_utils.utility import xmap_readers_mp
+from data_utils.augmentor.augmentation import AugmentationPipeline
+from data_utils.featurizer.speech_featurizer import SpeechFeaturizer
+from data_utils.speech import SpeechSegment
+from data_utils.normalizer import FeatureNormalizer
+
+
+class DataGenerator(object):
+ """
+ DataGenerator provides basic audio data preprocessing pipeline, and offers
+ data reader interfaces of PaddlePaddle requirements.
+
+ :param vocab_filepath: Vocabulary filepath for indexing tokenized
+ transcripts.
+ :type vocab_filepath: basestring
+ :param mean_std_filepath: File containing the pre-computed mean and stddev.
+ :type mean_std_filepath: None|basestring
+ :param augmentation_config: Augmentation configuration in json string.
+ Details see AugmentationPipeline.__doc__.
+ :type augmentation_config: str
+ :param max_duration: Audio with duration (in seconds) greater than
+ this will be discarded.
+ :type max_duration: float
+ :param min_duration: Audio with duration (in seconds) smaller than
+ this will be discarded.
+ :type min_duration: float
+ :param stride_ms: Striding size (in milliseconds) for generating frames.
+ :type stride_ms: float
+ :param window_ms: Window size (in milliseconds) for generating frames.
+ :type window_ms: float
+ :param max_freq: Used when specgram_type is 'linear', only FFT bins
+ corresponding to frequencies between [0, max_freq] are
+ returned.
+ :types max_freq: None|float
+ :param specgram_type: Specgram feature type. Options: 'linear'.
+ :type specgram_type: str
+ :param use_dB_normalization: Whether to normalize the audio to -20 dB
+ before extracting the features.
+ :type use_dB_normalization: bool
+ :param num_threads: Number of CPU threads for processing data.
+ :type num_threads: int
+ :param random_seed: Random seed.
+ :type random_seed: int
+ :param keep_transcription_text: If set to True, transcription text will
+ be passed forward directly without
+ converting to index sequence.
+ :type keep_transcription_text: bool
+ """
+
+ def __init__(self,
+ vocab_filepath,
+ mean_std_filepath,
+ augmentation_config='{}',
+ max_duration=float('inf'),
+ min_duration=0.0,
+ stride_ms=10.0,
+ window_ms=20.0,
+ max_freq=None,
+ specgram_type='linear',
+ use_dB_normalization=True,
+ num_threads=multiprocessing.cpu_count() // 2,
+ random_seed=0,
+ keep_transcription_text=False):
+ self._max_duration = max_duration
+ self._min_duration = min_duration
+ self._normalizer = FeatureNormalizer(mean_std_filepath)
+ self._augmentation_pipeline = AugmentationPipeline(
+ augmentation_config=augmentation_config, random_seed=random_seed)
+ self._speech_featurizer = SpeechFeaturizer(
+ vocab_filepath=vocab_filepath,
+ specgram_type=specgram_type,
+ stride_ms=stride_ms,
+ window_ms=window_ms,
+ max_freq=max_freq,
+ use_dB_normalization=use_dB_normalization)
+ self._num_threads = num_threads
+ self._rng = random.Random(random_seed)
+ self._keep_transcription_text = keep_transcription_text
+ self._epoch = 0
+ # for caching tar files info
+ self._local_data = local()
+ self._local_data.tar2info = {}
+ self._local_data.tar2object = {}
+
+ def process_utterance(self, audio_file, transcript):
+ """Load, augment, featurize and normalize for speech data.
+
+ :param audio_file: Filepath or file object of audio file.
+ :type audio_file: basestring | file
+ :param transcript: Transcription text.
+ :type transcript: basestring
+ :return: Tuple of audio feature tensor and data of transcription part,
+ where transcription part could be token ids or text.
+ :rtype: tuple of (2darray, list)
+ """
+ if isinstance(audio_file, basestring) and audio_file.startswith('tar:'):
+ speech_segment = SpeechSegment.from_file(
+ self._subfile_from_tar(audio_file), transcript)
+ else:
+ speech_segment = SpeechSegment.from_file(audio_file, transcript)
+ self._augmentation_pipeline.transform_audio(speech_segment)
+ specgram, transcript_part = self._speech_featurizer.featurize(
+ speech_segment, self._keep_transcription_text)
+ specgram = self._normalizer.apply(specgram)
+ return specgram, transcript_part
+
+ def batch_reader_creator(self,
+ manifest_path,
+ batch_size,
+ min_batch_size=1,
+ padding_to=-1,
+ flatten=False,
+ sortagrad=False,
+ shuffle_method="batch_shuffle"):
+ """
+ Batch data reader creator for audio data. Return a callable generator
+ function to produce batches of data.
+
+ Audio features within one batch will be padded with zeros to have the
+ same shape, or a user-defined shape.
+
+ :param manifest_path: Filepath of manifest for audio files.
+ :type manifest_path: basestring
+ :param batch_size: Number of instances in a batch.
+ :type batch_size: int
+ :param min_batch_size: Any batch with batch size smaller than this will
+ be discarded. (To be deprecated in the future.)
+ :type min_batch_size: int
+ :param padding_to: If set -1, the maximun shape in the batch
+ will be used as the target shape for padding.
+ Otherwise, `padding_to` will be the target shape.
+ :type padding_to: int
+ :param flatten: If set True, audio features will be flatten to 1darray.
+ :type flatten: bool
+ :param sortagrad: If set True, sort the instances by audio duration
+ in the first epoch for speed up training.
+ :type sortagrad: bool
+ :param shuffle_method: Shuffle method. Options:
+ '' or None: no shuffle.
+ 'instance_shuffle': instance-wise shuffle.
+ 'batch_shuffle': similarly-sized instances are
+ put into batches, and then
+ batch-wise shuffle the batches.
+ For more details, please see
+ ``_batch_shuffle.__doc__``.
+ 'batch_shuffle_clipped': 'batch_shuffle' with
+ head shift and tail
+ clipping. For more
+ details, please see
+ ``_batch_shuffle``.
+ If sortagrad is True, shuffle is disabled
+ for the first epoch.
+ :type shuffle_method: None|str
+ :return: Batch reader function, producing batches of data when called.
+ :rtype: callable
+ """
+
+ def batch_reader():
+ # read manifest
+ manifest = read_manifest(
+ manifest_path=manifest_path,
+ max_duration=self._max_duration,
+ min_duration=self._min_duration)
+ # sort (by duration) or batch-wise shuffle the manifest
+ if self._epoch == 0 and sortagrad:
+ manifest.sort(key=lambda x: x["duration"])
+ else:
+ if shuffle_method == "batch_shuffle":
+ manifest = self._batch_shuffle(
+ manifest, batch_size, clipped=False)
+ elif shuffle_method == "batch_shuffle_clipped":
+ manifest = self._batch_shuffle(
+ manifest, batch_size, clipped=True)
+ elif shuffle_method == "instance_shuffle":
+ self._rng.shuffle(manifest)
+ elif shuffle_method == None:
+ pass
+ else:
+ raise ValueError("Unknown shuffle method %s." %
+ shuffle_method)
+ # prepare batches
+ instance_reader, cleanup = self._instance_reader_creator(manifest)
+ batch = []
+ try:
+ for instance in instance_reader():
+ batch.append(instance)
+ if len(batch) == batch_size:
+ yield self._padding_batch(batch, padding_to, flatten)
+ batch = []
+ if len(batch) >= min_batch_size:
+ yield self._padding_batch(batch, padding_to, flatten)
+ finally:
+ cleanup()
+ self._epoch += 1
+
+ return batch_reader
+
+ @property
+ def feeding(self):
+ """Returns data reader's feeding dict.
+
+ :return: Data feeding dict.
+ :rtype: dict
+ """
+ feeding_dict = {"audio_spectrogram": 0, "transcript_text": 1}
+ return feeding_dict
+
+ @property
+ def vocab_size(self):
+ """Return the vocabulary size.
+
+ :return: Vocabulary size.
+ :rtype: int
+ """
+ return self._speech_featurizer.vocab_size
+
+ @property
+ def vocab_list(self):
+ """Return the vocabulary in list.
+
+ :return: Vocabulary in list.
+ :rtype: list
+ """
+ return self._speech_featurizer.vocab_list
+
+ def _parse_tar(self, file):
+ """Parse a tar file to get a tarfile object
+ and a map containing tarinfoes
+ """
+ result = {}
+ f = tarfile.open(file)
+ for tarinfo in f.getmembers():
+ result[tarinfo.name] = tarinfo
+ return f, result
+
+ def _subfile_from_tar(self, file):
+ """Get subfile object from tar.
+
+ It will return a subfile object from tar file
+ and cached tar file info for next reading request.
+ """
+ tarpath, filename = file.split(':', 1)[1].split('#', 1)
+ if 'tar2info' not in self._local_data.__dict__:
+ self._local_data.tar2info = {}
+ if 'tar2object' not in self._local_data.__dict__:
+ self._local_data.tar2object = {}
+ if tarpath not in self._local_data.tar2info:
+ object, infoes = self._parse_tar(tarpath)
+ self._local_data.tar2info[tarpath] = infoes
+ self._local_data.tar2object[tarpath] = object
+ return self._local_data.tar2object[tarpath].extractfile(
+ self._local_data.tar2info[tarpath][filename])
+
+ def _instance_reader_creator(self, manifest):
+ """
+ Instance reader creator. Create a callable function to produce
+ instances of data.
+
+ Instance: a tuple of ndarray of audio spectrogram and a list of
+ token indices for transcript.
+ """
+
+ def reader():
+ for instance in manifest:
+ yield instance
+
+ reader, cleanup_callback = xmap_readers_mp(
+ lambda instance: self.process_utterance(instance["audio_filepath"], instance["text"]),
+ reader, self._num_threads, 4096)
+
+ return reader, cleanup_callback
+
+ def _padding_batch(self, batch, padding_to=-1, flatten=False):
+ """
+ Padding audio features with zeros to make them have the same shape (or
+ a user-defined shape) within one bach.
+
+ If ``padding_to`` is -1, the maximun shape in the batch will be used
+ as the target shape for padding. Otherwise, `padding_to` will be the
+ target shape (only refers to the second axis).
+
+ If `flatten` is True, features will be flatten to 1darray.
+ """
+ new_batch = []
+ # get target shape
+ max_length = max([audio.shape[1] for audio, text in batch])
+ if padding_to != -1:
+ if padding_to < max_length:
+ raise ValueError("If padding_to is not -1, it should be larger "
+ "than any instance's shape in the batch")
+ max_length = padding_to
+ # padding
+ for audio, text in batch:
+ padded_audio = np.zeros([audio.shape[0], max_length])
+ padded_audio[:, :audio.shape[1]] = audio
+ if flatten:
+ padded_audio = padded_audio.flatten()
+ padded_instance = [padded_audio, text, audio.shape[1]]
+ new_batch.append(padded_instance)
+ return new_batch
+
+ def _batch_shuffle(self, manifest, batch_size, clipped=False):
+ """Put similarly-sized instances into minibatches for better efficiency
+ and make a batch-wise shuffle.
+
+ 1. Sort the audio clips by duration.
+ 2. Generate a random number `k`, k in [0, batch_size).
+ 3. Randomly shift `k` instances in order to create different batches
+ for different epochs. Create minibatches.
+ 4. Shuffle the minibatches.
+
+ :param manifest: Manifest contents. List of dict.
+ :type manifest: list
+ :param batch_size: Batch size. This size is also used for generate
+ a random number for batch shuffle.
+ :type batch_size: int
+ :param clipped: Whether to clip the heading (small shift) and trailing
+ (incomplete batch) instances.
+ :type clipped: bool
+ :return: Batch shuffled mainifest.
+ :rtype: list
+ """
+ manifest.sort(key=lambda x: x["duration"])
+ shift_len = self._rng.randint(0, batch_size - 1)
+ batch_manifest = zip(*[iter(manifest[shift_len:])] * batch_size)
+ self._rng.shuffle(batch_manifest)
+ batch_manifest = [item for batch in batch_manifest for item in batch]
+ if not clipped:
+ res_len = len(manifest) - shift_len - len(batch_manifest)
+ batch_manifest.extend(manifest[-res_len:])
+ batch_manifest.extend(manifest[0:shift_len])
+ return batch_manifest
diff --git a/data_utils/featurizer/__init__.py b/data_utils/featurizer/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/data_utils/featurizer/audio_featurizer.py b/data_utils/featurizer/audio_featurizer.py
new file mode 100644
index 000000000..0a54701bf
--- /dev/null
+++ b/data_utils/featurizer/audio_featurizer.py
@@ -0,0 +1,187 @@
+"""Contains the audio featurizer class."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+from data_utils.utility import read_manifest
+from data_utils.audio import AudioSegment
+from python_speech_features import mfcc
+from python_speech_features import delta
+
+
+class AudioFeaturizer(object):
+ """Audio featurizer, for extracting features from audio contents of
+ AudioSegment or SpeechSegment.
+
+ Currently, it supports feature types of linear spectrogram and mfcc.
+
+ :param specgram_type: Specgram feature type. Options: 'linear'.
+ :type specgram_type: str
+ :param stride_ms: Striding size (in milliseconds) for generating frames.
+ :type stride_ms: float
+ :param window_ms: Window size (in milliseconds) for generating frames.
+ :type window_ms: float
+ :param max_freq: When specgram_type is 'linear', only FFT bins
+ corresponding to frequencies between [0, max_freq] are
+ returned; when specgram_type is 'mfcc', max_feq is the
+ highest band edge of mel filters.
+ :types max_freq: None|float
+ :param target_sample_rate: Audio are resampled (if upsampling or
+ downsampling is allowed) to this before
+ extracting spectrogram features.
+ :type target_sample_rate: float
+ :param use_dB_normalization: Whether to normalize the audio to a certain
+ decibels before extracting the features.
+ :type use_dB_normalization: bool
+ :param target_dB: Target audio decibels for normalization.
+ :type target_dB: float
+ """
+
+ def __init__(self,
+ specgram_type='linear',
+ stride_ms=10.0,
+ window_ms=20.0,
+ max_freq=None,
+ target_sample_rate=16000,
+ use_dB_normalization=True,
+ target_dB=-20):
+ self._specgram_type = specgram_type
+ self._stride_ms = stride_ms
+ self._window_ms = window_ms
+ self._max_freq = max_freq
+ self._target_sample_rate = target_sample_rate
+ self._use_dB_normalization = use_dB_normalization
+ self._target_dB = target_dB
+
+ def featurize(self,
+ audio_segment,
+ allow_downsampling=True,
+ allow_upsampling=True):
+ """Extract audio features from AudioSegment or SpeechSegment.
+
+ :param audio_segment: Audio/speech segment to extract features from.
+ :type audio_segment: AudioSegment|SpeechSegment
+ :param allow_downsampling: Whether to allow audio downsampling before
+ featurizing.
+ :type allow_downsampling: bool
+ :param allow_upsampling: Whether to allow audio upsampling before
+ featurizing.
+ :type allow_upsampling: bool
+ :return: Spectrogram audio feature in 2darray.
+ :rtype: ndarray
+ :raises ValueError: If audio sample rate is not supported.
+ """
+ # upsampling or downsampling
+ if ((audio_segment.sample_rate > self._target_sample_rate and
+ allow_downsampling) or
+ (audio_segment.sample_rate < self._target_sample_rate and
+ allow_upsampling)):
+ audio_segment.resample(self._target_sample_rate)
+ if audio_segment.sample_rate != self._target_sample_rate:
+ raise ValueError("Audio sample rate is not supported. "
+ "Turn allow_downsampling or allow up_sampling on.")
+ # decibel normalization
+ if self._use_dB_normalization:
+ audio_segment.normalize(target_db=self._target_dB)
+ # extract spectrogram
+ return self._compute_specgram(audio_segment.samples,
+ audio_segment.sample_rate)
+
+ def _compute_specgram(self, samples, sample_rate):
+ """Extract various audio features."""
+ if self._specgram_type == 'linear':
+ return self._compute_linear_specgram(
+ samples, sample_rate, self._stride_ms, self._window_ms,
+ self._max_freq)
+ elif self._specgram_type == 'mfcc':
+ return self._compute_mfcc(samples, sample_rate, self._stride_ms,
+ self._window_ms, self._max_freq)
+ else:
+ raise ValueError("Unknown specgram_type %s. "
+ "Supported values: linear." % self._specgram_type)
+
+ def _compute_linear_specgram(self,
+ samples,
+ sample_rate,
+ stride_ms=10.0,
+ window_ms=20.0,
+ max_freq=None,
+ eps=1e-14):
+ """Compute the linear spectrogram from FFT energy."""
+ if max_freq is None:
+ max_freq = sample_rate / 2
+ if max_freq > sample_rate / 2:
+ raise ValueError("max_freq must not be greater than half of "
+ "sample rate.")
+ if stride_ms > window_ms:
+ raise ValueError("Stride size must not be greater than "
+ "window size.")
+ stride_size = int(0.001 * sample_rate * stride_ms)
+ window_size = int(0.001 * sample_rate * window_ms)
+ specgram, freqs = self._specgram_real(
+ samples,
+ window_size=window_size,
+ stride_size=stride_size,
+ sample_rate=sample_rate)
+ ind = np.where(freqs <= max_freq)[0][-1] + 1
+ return np.log(specgram[:ind, :] + eps)
+
+ def _specgram_real(self, samples, window_size, stride_size, sample_rate):
+ """Compute the spectrogram for samples from a real signal."""
+ # extract strided windows
+ truncate_size = (len(samples) - window_size) % stride_size
+ samples = samples[:len(samples) - truncate_size]
+ nshape = (window_size, (len(samples) - window_size) // stride_size + 1)
+ nstrides = (samples.strides[0], samples.strides[0] * stride_size)
+ windows = np.lib.stride_tricks.as_strided(
+ samples, shape=nshape, strides=nstrides)
+ assert np.all(
+ windows[:, 1] == samples[stride_size:(stride_size + window_size)])
+ # window weighting, squared Fast Fourier Transform (fft), scaling
+ weighting = np.hanning(window_size)[:, None]
+ fft = np.fft.rfft(windows * weighting, axis=0)
+ fft = np.absolute(fft)
+ fft = fft**2
+ scale = np.sum(weighting**2) * sample_rate
+ fft[1:-1, :] *= (2.0 / scale)
+ fft[(0, -1), :] /= scale
+ # prepare fft frequency list
+ freqs = float(sample_rate) / window_size * np.arange(fft.shape[0])
+ return fft, freqs
+
+ def _compute_mfcc(self,
+ samples,
+ sample_rate,
+ stride_ms=10.0,
+ window_ms=20.0,
+ max_freq=None):
+ """Compute mfcc from samples."""
+ if max_freq is None:
+ max_freq = sample_rate / 2
+ if max_freq > sample_rate / 2:
+ raise ValueError("max_freq must not be greater than half of "
+ "sample rate.")
+ if stride_ms > window_ms:
+ raise ValueError("Stride size must not be greater than "
+ "window size.")
+ # compute the 13 cepstral coefficients, and the first one is replaced
+ # by log(frame energy)
+ mfcc_feat = mfcc(
+ signal=samples,
+ samplerate=sample_rate,
+ winlen=0.001 * window_ms,
+ winstep=0.001 * stride_ms,
+ highfreq=max_freq)
+ # Deltas
+ d_mfcc_feat = delta(mfcc_feat, 2)
+ # Deltas-Deltas
+ dd_mfcc_feat = delta(d_mfcc_feat, 2)
+ # transpose
+ mfcc_feat = np.transpose(mfcc_feat)
+ d_mfcc_feat = np.transpose(d_mfcc_feat)
+ dd_mfcc_feat = np.transpose(dd_mfcc_feat)
+ # concat above three features
+ concat_mfcc_feat = np.concatenate(
+ (mfcc_feat, d_mfcc_feat, dd_mfcc_feat))
+ return concat_mfcc_feat
diff --git a/data_utils/featurizer/speech_featurizer.py b/data_utils/featurizer/speech_featurizer.py
new file mode 100644
index 000000000..4555dc31d
--- /dev/null
+++ b/data_utils/featurizer/speech_featurizer.py
@@ -0,0 +1,98 @@
+"""Contains the speech featurizer class."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from data_utils.featurizer.audio_featurizer import AudioFeaturizer
+from data_utils.featurizer.text_featurizer import TextFeaturizer
+
+
+class SpeechFeaturizer(object):
+ """Speech featurizer, for extracting features from both audio and transcript
+ contents of SpeechSegment.
+
+ Currently, for audio parts, it supports feature types of linear
+ spectrogram and mfcc; for transcript parts, it only supports char-level
+ tokenizing and conversion into a list of token indices. Note that the
+ token indexing order follows the given vocabulary file.
+
+ :param vocab_filepath: Filepath to load vocabulary for token indices
+ conversion.
+ :type specgram_type: basestring
+ :param specgram_type: Specgram feature type. Options: 'linear', 'mfcc'.
+ :type specgram_type: str
+ :param stride_ms: Striding size (in milliseconds) for generating frames.
+ :type stride_ms: float
+ :param window_ms: Window size (in milliseconds) for generating frames.
+ :type window_ms: float
+ :param max_freq: When specgram_type is 'linear', only FFT bins
+ corresponding to frequencies between [0, max_freq] are
+ returned; when specgram_type is 'mfcc', max_freq is the
+ highest band edge of mel filters.
+ :types max_freq: None|float
+ :param target_sample_rate: Speech are resampled (if upsampling or
+ downsampling is allowed) to this before
+ extracting spectrogram features.
+ :type target_sample_rate: float
+ :param use_dB_normalization: Whether to normalize the audio to a certain
+ decibels before extracting the features.
+ :type use_dB_normalization: bool
+ :param target_dB: Target audio decibels for normalization.
+ :type target_dB: float
+ """
+
+ def __init__(self,
+ vocab_filepath,
+ specgram_type='linear',
+ stride_ms=10.0,
+ window_ms=20.0,
+ max_freq=None,
+ target_sample_rate=16000,
+ use_dB_normalization=True,
+ target_dB=-20):
+ self._audio_featurizer = AudioFeaturizer(
+ specgram_type=specgram_type,
+ stride_ms=stride_ms,
+ window_ms=window_ms,
+ max_freq=max_freq,
+ target_sample_rate=target_sample_rate,
+ use_dB_normalization=use_dB_normalization,
+ target_dB=target_dB)
+ self._text_featurizer = TextFeaturizer(vocab_filepath)
+
+ def featurize(self, speech_segment, keep_transcription_text):
+ """Extract features for speech segment.
+
+ 1. For audio parts, extract the audio features.
+ 2. For transcript parts, keep the original text or convert text string
+ to a list of token indices in char-level.
+
+ :param audio_segment: Speech segment to extract features from.
+ :type audio_segment: SpeechSegment
+ :return: A tuple of 1) spectrogram audio feature in 2darray, 2) list of
+ char-level token indices.
+ :rtype: tuple
+ """
+ audio_feature = self._audio_featurizer.featurize(speech_segment)
+ if keep_transcription_text:
+ return audio_feature, speech_segment.transcript
+ text_ids = self._text_featurizer.featurize(speech_segment.transcript)
+ return audio_feature, text_ids
+
+ @property
+ def vocab_size(self):
+ """Return the vocabulary size.
+
+ :return: Vocabulary size.
+ :rtype: int
+ """
+ return self._text_featurizer.vocab_size
+
+ @property
+ def vocab_list(self):
+ """Return the vocabulary in list.
+
+ :return: Vocabulary in list.
+ :rtype: list
+ """
+ return self._text_featurizer.vocab_list
diff --git a/data_utils/featurizer/text_featurizer.py b/data_utils/featurizer/text_featurizer.py
new file mode 100644
index 000000000..89202163c
--- /dev/null
+++ b/data_utils/featurizer/text_featurizer.py
@@ -0,0 +1,68 @@
+"""Contains the text featurizer class."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import codecs
+
+
+class TextFeaturizer(object):
+ """Text featurizer, for processing or extracting features from text.
+
+ Currently, it only supports char-level tokenizing and conversion into
+ a list of token indices. Note that the token indexing order follows the
+ given vocabulary file.
+
+ :param vocab_filepath: Filepath to load vocabulary for token indices
+ conversion.
+ :type specgram_type: basestring
+ """
+
+ def __init__(self, vocab_filepath):
+ self._vocab_dict, self._vocab_list = self._load_vocabulary_from_file(
+ vocab_filepath)
+
+ def featurize(self, text):
+ """Convert text string to a list of token indices in char-level.Note
+ that the token indexing order follows the given vocabulary file.
+
+ :param text: Text to process.
+ :type text: basestring
+ :return: List of char-level token indices.
+ :rtype: list
+ """
+ tokens = self._char_tokenize(text)
+ return [self._vocab_dict[token] for token in tokens]
+
+ @property
+ def vocab_size(self):
+ """Return the vocabulary size.
+
+ :return: Vocabulary size.
+ :rtype: int
+ """
+ return len(self._vocab_list)
+
+ @property
+ def vocab_list(self):
+ """Return the vocabulary in list.
+
+ :return: Vocabulary in list.
+ :rtype: list
+ """
+ return self._vocab_list
+
+ def _char_tokenize(self, text):
+ """Character tokenizer."""
+ return list(text.strip())
+
+ def _load_vocabulary_from_file(self, vocab_filepath):
+ """Load vocabulary from file."""
+ vocab_lines = []
+ with codecs.open(vocab_filepath, 'r', 'utf-8') as file:
+ vocab_lines.extend(file.readlines())
+ vocab_list = [line[:-1] for line in vocab_lines]
+ vocab_dict = dict(
+ [(token, id) for (id, token) in enumerate(vocab_list)])
+ return vocab_dict, vocab_list
diff --git a/data_utils/normalizer.py b/data_utils/normalizer.py
new file mode 100644
index 000000000..7c2e05c9d
--- /dev/null
+++ b/data_utils/normalizer.py
@@ -0,0 +1,87 @@
+"""Contains feature normalizers."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import random
+from data_utils.utility import read_manifest
+from data_utils.audio import AudioSegment
+
+
+class FeatureNormalizer(object):
+ """Feature normalizer. Normalize features to be of zero mean and unit
+ stddev.
+
+ if mean_std_filepath is provided (not None), the normalizer will directly
+ initilize from the file. Otherwise, both manifest_path and featurize_func
+ should be given for on-the-fly mean and stddev computing.
+
+ :param mean_std_filepath: File containing the pre-computed mean and stddev.
+ :type mean_std_filepath: None|basestring
+ :param manifest_path: Manifest of instances for computing mean and stddev.
+ :type meanifest_path: None|basestring
+ :param featurize_func: Function to extract features. It should be callable
+ with ``featurize_func(audio_segment)``.
+ :type featurize_func: None|callable
+ :param num_samples: Number of random samples for computing mean and stddev.
+ :type num_samples: int
+ :param random_seed: Random seed for sampling instances.
+ :type random_seed: int
+ :raises ValueError: If both mean_std_filepath and manifest_path
+ (or both mean_std_filepath and featurize_func) are None.
+ """
+
+ def __init__(self,
+ mean_std_filepath,
+ manifest_path=None,
+ featurize_func=None,
+ num_samples=500,
+ random_seed=0):
+ if not mean_std_filepath:
+ if not (manifest_path and featurize_func):
+ raise ValueError("If mean_std_filepath is None, meanifest_path "
+ "and featurize_func should not be None.")
+ self._rng = random.Random(random_seed)
+ self._compute_mean_std(manifest_path, featurize_func, num_samples)
+ else:
+ self._read_mean_std_from_file(mean_std_filepath)
+
+ def apply(self, features, eps=1e-14):
+ """Normalize features to be of zero mean and unit stddev.
+
+ :param features: Input features to be normalized.
+ :type features: ndarray
+ :param eps: added to stddev to provide numerical stablibity.
+ :type eps: float
+ :return: Normalized features.
+ :rtype: ndarray
+ """
+ return (features - self._mean) / (self._std + eps)
+
+ def write_to_file(self, filepath):
+ """Write the mean and stddev to the file.
+
+ :param filepath: File to write mean and stddev.
+ :type filepath: basestring
+ """
+ np.savez(filepath, mean=self._mean, std=self._std)
+
+ def _read_mean_std_from_file(self, filepath):
+ """Load mean and std from file."""
+ npzfile = np.load(filepath)
+ self._mean = npzfile["mean"]
+ self._std = npzfile["std"]
+
+ def _compute_mean_std(self, manifest_path, featurize_func, num_samples):
+ """Compute mean and std from randomly sampled instances."""
+ manifest = read_manifest(manifest_path)
+ sampled_manifest = self._rng.sample(manifest, num_samples)
+ features = []
+ for instance in sampled_manifest:
+ features.append(
+ featurize_func(
+ AudioSegment.from_file(instance["audio_filepath"])))
+ features = np.hstack(features)
+ self._mean = np.mean(features, axis=1).reshape([-1, 1])
+ self._std = np.std(features, axis=1).reshape([-1, 1])
diff --git a/data_utils/speech.py b/data_utils/speech.py
new file mode 100644
index 000000000..0cea88730
--- /dev/null
+++ b/data_utils/speech.py
@@ -0,0 +1,143 @@
+"""Contains the speech segment class."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from data_utils.audio import AudioSegment
+
+
+class SpeechSegment(AudioSegment):
+ """Speech segment abstraction, a subclass of AudioSegment,
+ with an additional transcript.
+
+ :param samples: Audio samples [num_samples x num_channels].
+ :type samples: ndarray.float32
+ :param sample_rate: Audio sample rate.
+ :type sample_rate: int
+ :param transcript: Transcript text for the speech.
+ :type transript: basestring
+ :raises TypeError: If the sample data type is not float or int.
+ """
+
+ def __init__(self, samples, sample_rate, transcript):
+ AudioSegment.__init__(self, samples, sample_rate)
+ self._transcript = transcript
+
+ def __eq__(self, other):
+ """Return whether two objects are equal.
+ """
+ if not AudioSegment.__eq__(self, other):
+ return False
+ if self._transcript != other._transcript:
+ return False
+ return True
+
+ def __ne__(self, other):
+ """Return whether two objects are unequal."""
+ return not self.__eq__(other)
+
+ @classmethod
+ def from_file(cls, filepath, transcript):
+ """Create speech segment from audio file and corresponding transcript.
+
+ :param filepath: Filepath or file object to audio file.
+ :type filepath: basestring|file
+ :param transcript: Transcript text for the speech.
+ :type transript: basestring
+ :return: Speech segment instance.
+ :rtype: SpeechSegment
+ """
+ audio = AudioSegment.from_file(filepath)
+ return cls(audio.samples, audio.sample_rate, transcript)
+
+ @classmethod
+ def from_bytes(cls, bytes, transcript):
+ """Create speech segment from a byte string and corresponding
+ transcript.
+
+ :param bytes: Byte string containing audio samples.
+ :type bytes: str
+ :param transcript: Transcript text for the speech.
+ :type transript: basestring
+ :return: Speech segment instance.
+ :rtype: Speech Segment
+ """
+ audio = AudioSegment.from_bytes(bytes)
+ return cls(audio.samples, audio.sample_rate, transcript)
+
+ @classmethod
+ def concatenate(cls, *segments):
+ """Concatenate an arbitrary number of speech segments together, both
+ audio and transcript will be concatenated.
+
+ :param *segments: Input speech segments to be concatenated.
+ :type *segments: tuple of SpeechSegment
+ :return: Speech segment instance.
+ :rtype: SpeechSegment
+ :raises ValueError: If the number of segments is zero, or if the
+ sample_rate of any two segments does not match.
+ :raises TypeError: If any segment is not SpeechSegment instance.
+ """
+ if len(segments) == 0:
+ raise ValueError("No speech segments are given to concatenate.")
+ sample_rate = segments[0]._sample_rate
+ transcripts = ""
+ for seg in segments:
+ if sample_rate != seg._sample_rate:
+ raise ValueError("Can't concatenate segments with "
+ "different sample rates")
+ if type(seg) is not cls:
+ raise TypeError("Only speech segments of the same type "
+ "instance can be concatenated.")
+ transcripts += seg._transcript
+ samples = np.concatenate([seg.samples for seg in segments])
+ return cls(samples, sample_rate, transcripts)
+
+ @classmethod
+ def slice_from_file(cls, filepath, transcript, start=None, end=None):
+ """Loads a small section of an speech without having to load
+ the entire file into the memory which can be incredibly wasteful.
+
+ :param filepath: Filepath or file object to audio file.
+ :type filepath: basestring|file
+ :param start: Start time in seconds. If start is negative, it wraps
+ around from the end. If not provided, this function
+ reads from the very beginning.
+ :type start: float
+ :param end: End time in seconds. If end is negative, it wraps around
+ from the end. If not provided, the default behvaior is
+ to read to the end of the file.
+ :type end: float
+ :param transcript: Transcript text for the speech. if not provided,
+ the defaults is an empty string.
+ :type transript: basestring
+ :return: SpeechSegment instance of the specified slice of the input
+ speech file.
+ :rtype: SpeechSegment
+ """
+ audio = AudioSegment.slice_from_file(filepath, start, end)
+ return cls(audio.samples, audio.sample_rate, transcript)
+
+ @classmethod
+ def make_silence(cls, duration, sample_rate):
+ """Creates a silent speech segment of the given duration and
+ sample rate, transcript will be an empty string.
+
+ :param duration: Length of silence in seconds.
+ :type duration: float
+ :param sample_rate: Sample rate.
+ :type sample_rate: float
+ :return: Silence of the given duration.
+ :rtype: SpeechSegment
+ """
+ audio = AudioSegment.make_silence(duration, sample_rate)
+ return cls(audio.samples, audio.sample_rate, "")
+
+ @property
+ def transcript(self):
+ """Return the transcript text.
+
+ :return: Transcript text for the speech.
+ :rtype: basestring
+ """
+ return self._transcript
diff --git a/data_utils/utility.py b/data_utils/utility.py
new file mode 100644
index 000000000..89a74c41a
--- /dev/null
+++ b/data_utils/utility.py
@@ -0,0 +1,214 @@
+"""Contains data helper functions."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import json
+import codecs
+import os
+import tarfile
+import time
+from Queue import Queue
+from threading import Thread
+from multiprocessing import Process, Manager, Value
+from paddle.v2.dataset.common import md5file
+
+
+def read_manifest(manifest_path, max_duration=float('inf'), min_duration=0.0):
+ """Load and parse manifest file.
+
+ Instances with durations outside [min_duration, max_duration] will be
+ filtered out.
+
+ :param manifest_path: Manifest file to load and parse.
+ :type manifest_path: basestring
+ :param max_duration: Maximal duration in seconds for instance filter.
+ :type max_duration: float
+ :param min_duration: Minimal duration in seconds for instance filter.
+ :type min_duration: float
+ :return: Manifest parsing results. List of dict.
+ :rtype: list
+ :raises IOError: If failed to parse the manifest.
+ """
+ manifest = []
+ for json_line in codecs.open(manifest_path, 'r', 'utf-8'):
+ try:
+ json_data = json.loads(json_line)
+ except Exception as e:
+ raise IOError("Error reading manifest: %s" % str(e))
+ if (json_data["duration"] <= max_duration and
+ json_data["duration"] >= min_duration):
+ manifest.append(json_data)
+ return manifest
+
+
+def getfile_insensitive(path):
+ """Get the actual file path when given insensitive filename."""
+ directory, filename = os.path.split(path)
+ directory, filename = (directory or '.'), filename.lower()
+ for f in os.listdir(directory):
+ newpath = os.path.join(directory, f)
+ if os.path.isfile(newpath) and f.lower() == filename:
+ return newpath
+
+
+def download_multi(url, target_dir, extra_args):
+ """Download multiple files from url to target_dir."""
+ if not os.path.exists(target_dir): os.makedirs(target_dir)
+ print("Downloading %s ..." % url)
+ ret_code = os.system("wget -c " + url + ' ' + extra_args + " -P " +
+ target_dir)
+ return ret_code
+
+
+def download(url, md5sum, target_dir):
+ """Download file from url to target_dir, and check md5sum."""
+ if not os.path.exists(target_dir): os.makedirs(target_dir)
+ filepath = os.path.join(target_dir, url.split("/")[-1])
+ if not (os.path.exists(filepath) and md5file(filepath) == md5sum):
+ print("Downloading %s ..." % url)
+ os.system("wget -c " + url + " -P " + target_dir)
+ print("\nMD5 Chesksum %s ..." % filepath)
+ if not md5file(filepath) == md5sum:
+ raise RuntimeError("MD5 checksum failed.")
+ else:
+ print("File exists, skip downloading. (%s)" % filepath)
+ return filepath
+
+
+def unpack(filepath, target_dir, rm_tar=False):
+ """Unpack the file to the target_dir."""
+ print("Unpacking %s ..." % filepath)
+ tar = tarfile.open(filepath)
+ tar.extractall(target_dir)
+ tar.close()
+ if rm_tar == True:
+ os.remove(filepath)
+
+
+class XmapEndSignal():
+ pass
+
+
+def xmap_readers_mp(mapper, reader, process_num, buffer_size, order=False):
+ """A multiprocessing pipeline wrapper for the data reader.
+
+ :param mapper: Function to map sample.
+ :type mapper: callable
+ :param reader: Given data reader.
+ :type reader: callable
+ :param process_num: Number of processes in the pipeline
+ :type process_num: int
+ :param buffer_size: Maximal buffer size.
+ :type buffer_size: int
+ :return: The wrappered reader and cleanup callback
+ :rtype: tuple
+ """
+ end_flag = XmapEndSignal()
+
+ read_workers = []
+ handle_workers = []
+ flush_workers = []
+
+ read_exit_flag = Value('i', 0)
+ handle_exit_flag = Value('i', 0)
+ flush_exit_flag = Value('i', 0)
+
+ # define a worker to read samples from reader to in_queue with order flag
+ def order_read_worker(reader, in_queue):
+ for order_id, sample in enumerate(reader()):
+ if read_exit_flag.value == 1: break
+ in_queue.put((order_id, sample))
+ in_queue.put(end_flag)
+ # the reading worker should not exit until all handling work exited
+ while handle_exit_flag.value == 0 or read_exit_flag.value == 0:
+ time.sleep(0.001)
+
+ # define a worker to handle samples from in_queue by mapper and put results
+ # to out_queue with order
+ def order_handle_worker(in_queue, out_queue, mapper, out_order):
+ ins = in_queue.get()
+ while not isinstance(ins, XmapEndSignal):
+ if handle_exit_flag.value == 1: break
+ order_id, sample = ins
+ result = mapper(sample)
+ while order_id != out_order[0]:
+ time.sleep(0.001)
+ out_queue.put(result)
+ out_order[0] += 1
+ ins = in_queue.get()
+ in_queue.put(end_flag)
+ out_queue.put(end_flag)
+ # wait for exit of flushing worker
+ while flush_exit_flag.value == 0 or handle_exit_flag.value == 0:
+ time.sleep(0.001)
+ read_exit_flag.value = 1
+ handle_exit_flag.value = 1
+
+ # define a thread worker to flush samples from Manager.Queue to Queue
+ # for acceleration
+ def flush_worker(in_queue, out_queue):
+ finish = 0
+ while finish < process_num and flush_exit_flag.value == 0:
+ sample = in_queue.get()
+ if isinstance(sample, XmapEndSignal):
+ finish += 1
+ else:
+ out_queue.put(sample)
+ out_queue.put(end_flag)
+ handle_exit_flag.value = 1
+ flush_exit_flag.value = 1
+
+ def cleanup():
+ # first exit flushing workers
+ flush_exit_flag.value = 1
+ for w in flush_workers:
+ w.join()
+ # next exit handling workers
+ handle_exit_flag.value = 1
+ for w in handle_workers:
+ w.join()
+ # last exit reading workers
+ read_exit_flag.value = 1
+ for w in read_workers:
+ w.join()
+
+ def xreader():
+ # prepare shared memory
+ manager = Manager()
+ in_queue = manager.Queue(buffer_size)
+ out_queue = manager.Queue(buffer_size)
+ out_order = manager.list([0])
+
+ # start a read worker in a process
+ target = order_read_worker
+ p = Process(target=target, args=(reader, in_queue))
+ p.daemon = True
+ p.start()
+ read_workers.append(p)
+
+ # start handle_workers with multiple processes
+ target = order_handle_worker
+ args = (in_queue, out_queue, mapper, out_order)
+ workers = [
+ Process(target=target, args=args) for _ in xrange(process_num)
+ ]
+ for w in workers:
+ w.daemon = True
+ w.start()
+ handle_workers.append(w)
+
+ # start a thread to read data from slow Manager.Queue
+ flush_queue = Queue(buffer_size)
+ t = Thread(target=flush_worker, args=(out_queue, flush_queue))
+ t.daemon = True
+ t.start()
+ flush_workers.append(t)
+
+ # get results
+ sample = flush_queue.get()
+ while not isinstance(sample, XmapEndSignal):
+ yield sample
+ sample = flush_queue.get()
+
+ return xreader, cleanup
diff --git a/decoders/__init__.py b/decoders/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/decoders/decoders_deprecated.py b/decoders/decoders_deprecated.py
new file mode 100644
index 000000000..17b28b0d0
--- /dev/null
+++ b/decoders/decoders_deprecated.py
@@ -0,0 +1,238 @@
+"""Contains various CTC decoders."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from itertools import groupby
+import numpy as np
+from math import log
+import multiprocessing
+
+
+def ctc_greedy_decoder(probs_seq, vocabulary):
+ """CTC greedy (best path) decoder.
+
+ Path consisting of the most probable tokens are further post-processed to
+ remove consecutive repetitions and all blanks.
+
+ :param probs_seq: 2-D list of probabilities over the vocabulary for each
+ character. Each element is a list of float probabilities
+ for one character.
+ :type probs_seq: list
+ :param vocabulary: Vocabulary list.
+ :type vocabulary: list
+ :return: Decoding result string.
+ :rtype: baseline
+ """
+ # dimension verification
+ for probs in probs_seq:
+ if not len(probs) == len(vocabulary) + 1:
+ raise ValueError("probs_seq dimension mismatchedd with vocabulary")
+ # argmax to get the best index for each time step
+ max_index_list = list(np.array(probs_seq).argmax(axis=1))
+ # remove consecutive duplicate indexes
+ index_list = [index_group[0] for index_group in groupby(max_index_list)]
+ # remove blank indexes
+ blank_index = len(vocabulary)
+ index_list = [index for index in index_list if index != blank_index]
+ # convert index list to string
+ return ''.join([vocabulary[index] for index in index_list])
+
+
+def ctc_beam_search_decoder(probs_seq,
+ beam_size,
+ vocabulary,
+ cutoff_prob=1.0,
+ cutoff_top_n=40,
+ ext_scoring_func=None,
+ nproc=False):
+ """CTC Beam search decoder.
+
+ It utilizes beam search to approximately select top best decoding
+ labels and returning results in the descending order.
+ The implementation is based on Prefix Beam Search
+ (https://arxiv.org/abs/1408.2873), and the unclear part is
+ redesigned. Two important modifications: 1) in the iterative computation
+ of probabilities, the assignment operation is changed to accumulation for
+ one prefix may comes from different paths; 2) the if condition "if l^+ not
+ in A_prev then" after probabilities' computation is deprecated for it is
+ hard to understand and seems unnecessary.
+
+ :param probs_seq: 2-D list of probability distributions over each time
+ step, with each element being a list of normalized
+ probabilities over vocabulary and blank.
+ :type probs_seq: 2-D list
+ :param beam_size: Width for beam search.
+ :type beam_size: int
+ :param vocabulary: Vocabulary list.
+ :type vocabulary: list
+ :param cutoff_prob: Cutoff probability in pruning,
+ default 1.0, no pruning.
+ :type cutoff_prob: float
+ :param ext_scoring_func: External scoring function for
+ partially decoded sentence, e.g. word count
+ or language model.
+ :type external_scoring_func: callable
+ :param nproc: Whether the decoder used in multiprocesses.
+ :type nproc: bool
+ :return: List of tuples of log probability and sentence as decoding
+ results, in descending order of the probability.
+ :rtype: list
+ """
+ # dimension check
+ for prob_list in probs_seq:
+ if not len(prob_list) == len(vocabulary) + 1:
+ raise ValueError("The shape of prob_seq does not match with the "
+ "shape of the vocabulary.")
+
+ # blank_id assign
+ blank_id = len(vocabulary)
+
+ # If the decoder called in the multiprocesses, then use the global scorer
+ # instantiated in ctc_beam_search_decoder_batch().
+ if nproc is True:
+ global ext_nproc_scorer
+ ext_scoring_func = ext_nproc_scorer
+
+ ## initialize
+ # prefix_set_prev: the set containing selected prefixes
+ # probs_b_prev: prefixes' probability ending with blank in previous step
+ # probs_nb_prev: prefixes' probability ending with non-blank in previous step
+ prefix_set_prev = {'\t': 1.0}
+ probs_b_prev, probs_nb_prev = {'\t': 1.0}, {'\t': 0.0}
+
+ ## extend prefix in loop
+ for time_step in xrange(len(probs_seq)):
+ # prefix_set_next: the set containing candidate prefixes
+ # probs_b_cur: prefixes' probability ending with blank in current step
+ # probs_nb_cur: prefixes' probability ending with non-blank in current step
+ prefix_set_next, probs_b_cur, probs_nb_cur = {}, {}, {}
+
+ prob_idx = list(enumerate(probs_seq[time_step]))
+ cutoff_len = len(prob_idx)
+ #If pruning is enabled
+ if cutoff_prob < 1.0 or cutoff_top_n < cutoff_len:
+ prob_idx = sorted(prob_idx, key=lambda asd: asd[1], reverse=True)
+ cutoff_len, cum_prob = 0, 0.0
+ for i in xrange(len(prob_idx)):
+ cum_prob += prob_idx[i][1]
+ cutoff_len += 1
+ if cum_prob >= cutoff_prob:
+ break
+ cutoff_len = min(cutoff_len, cutoff_top_n)
+ prob_idx = prob_idx[0:cutoff_len]
+
+ for l in prefix_set_prev:
+ if not prefix_set_next.has_key(l):
+ probs_b_cur[l], probs_nb_cur[l] = 0.0, 0.0
+
+ # extend prefix by travering prob_idx
+ for index in xrange(cutoff_len):
+ c, prob_c = prob_idx[index][0], prob_idx[index][1]
+
+ if c == blank_id:
+ probs_b_cur[l] += prob_c * (
+ probs_b_prev[l] + probs_nb_prev[l])
+ else:
+ last_char = l[-1]
+ new_char = vocabulary[c]
+ l_plus = l + new_char
+ if not prefix_set_next.has_key(l_plus):
+ probs_b_cur[l_plus], probs_nb_cur[l_plus] = 0.0, 0.0
+
+ if new_char == last_char:
+ probs_nb_cur[l_plus] += prob_c * probs_b_prev[l]
+ probs_nb_cur[l] += prob_c * probs_nb_prev[l]
+ elif new_char == ' ':
+ if (ext_scoring_func is None) or (len(l) == 1):
+ score = 1.0
+ else:
+ prefix = l[1:]
+ score = ext_scoring_func(prefix)
+ probs_nb_cur[l_plus] += score * prob_c * (
+ probs_b_prev[l] + probs_nb_prev[l])
+ else:
+ probs_nb_cur[l_plus] += prob_c * (
+ probs_b_prev[l] + probs_nb_prev[l])
+ # add l_plus into prefix_set_next
+ prefix_set_next[l_plus] = probs_nb_cur[
+ l_plus] + probs_b_cur[l_plus]
+ # add l into prefix_set_next
+ prefix_set_next[l] = probs_b_cur[l] + probs_nb_cur[l]
+ # update probs
+ probs_b_prev, probs_nb_prev = probs_b_cur, probs_nb_cur
+
+ ## store top beam_size prefixes
+ prefix_set_prev = sorted(
+ prefix_set_next.iteritems(), key=lambda asd: asd[1], reverse=True)
+ if beam_size < len(prefix_set_prev):
+ prefix_set_prev = prefix_set_prev[:beam_size]
+ prefix_set_prev = dict(prefix_set_prev)
+
+ beam_result = []
+ for seq, prob in prefix_set_prev.items():
+ if prob > 0.0 and len(seq) > 1:
+ result = seq[1:]
+ # score last word by external scorer
+ if (ext_scoring_func is not None) and (result[-1] != ' '):
+ prob = prob * ext_scoring_func(result)
+ log_prob = log(prob)
+ beam_result.append((log_prob, result))
+ else:
+ beam_result.append((float('-inf'), ''))
+
+ ## output top beam_size decoding results
+ beam_result = sorted(beam_result, key=lambda asd: asd[0], reverse=True)
+ return beam_result
+
+
+def ctc_beam_search_decoder_batch(probs_split,
+ beam_size,
+ vocabulary,
+ num_processes,
+ cutoff_prob=1.0,
+ cutoff_top_n=40,
+ ext_scoring_func=None):
+ """CTC beam search decoder using multiple processes.
+
+ :param probs_seq: 3-D list with each element as an instance of 2-D list
+ of probabilities used by ctc_beam_search_decoder().
+ :type probs_seq: 3-D list
+ :param beam_size: Width for beam search.
+ :type beam_size: int
+ :param vocabulary: Vocabulary list.
+ :type vocabulary: list
+ :param num_processes: Number of parallel processes.
+ :type num_processes: int
+ :param cutoff_prob: Cutoff probability in pruning,
+ default 1.0, no pruning.
+ :type cutoff_prob: float
+ :param num_processes: Number of parallel processes.
+ :type num_processes: int
+ :param ext_scoring_func: External scoring function for
+ partially decoded sentence, e.g. word count
+ or language model.
+ :type external_scoring_function: callable
+ :return: List of tuples of log probability and sentence as decoding
+ results, in descending order of the probability.
+ :rtype: list
+ """
+ if not num_processes > 0:
+ raise ValueError("Number of processes must be positive!")
+
+ # use global variable to pass the externnal scorer to beam search decoder
+ global ext_nproc_scorer
+ ext_nproc_scorer = ext_scoring_func
+ nproc = True
+
+ pool = multiprocessing.Pool(processes=num_processes)
+ results = []
+ for i, probs_list in enumerate(probs_split):
+ args = (probs_list, beam_size, vocabulary, cutoff_prob, cutoff_top_n,
+ None, nproc)
+ results.append(pool.apply_async(ctc_beam_search_decoder, args))
+
+ pool.close()
+ pool.join()
+ beam_search_results = [result.get() for result in results]
+ return beam_search_results
diff --git a/decoders/scorer_deprecated.py b/decoders/scorer_deprecated.py
new file mode 100644
index 000000000..c6a661030
--- /dev/null
+++ b/decoders/scorer_deprecated.py
@@ -0,0 +1,68 @@
+"""External Scorer for Beam Search Decoder."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import kenlm
+import numpy as np
+
+
+class Scorer(object):
+ """External scorer to evaluate a prefix or whole sentence in
+ beam search decoding, including the score from n-gram language
+ model and word count.
+
+ :param alpha: Parameter associated with language model. Don't use
+ language model when alpha = 0.
+ :type alpha: float
+ :param beta: Parameter associated with word count. Don't use word
+ count when beta = 0.
+ :type beta: float
+ :model_path: Path to load language model.
+ :type model_path: basestring
+ """
+
+ def __init__(self, alpha, beta, model_path):
+ self._alpha = alpha
+ self._beta = beta
+ if not os.path.isfile(model_path):
+ raise IOError("Invaid language model path: %s" % model_path)
+ self._language_model = kenlm.LanguageModel(model_path)
+
+ # n-gram language model scoring
+ def _language_model_score(self, sentence):
+ #log10 prob of last word
+ log_cond_prob = list(
+ self._language_model.full_scores(sentence, eos=False))[-1][0]
+ return np.power(10, log_cond_prob)
+
+ # word insertion term
+ def _word_count(self, sentence):
+ words = sentence.strip().split(' ')
+ return len(words)
+
+ # reset alpha and beta
+ def reset_params(self, alpha, beta):
+ self._alpha = alpha
+ self._beta = beta
+
+ # execute evaluation
+ def __call__(self, sentence, log=False):
+ """Evaluation function, gathering all the different scores
+ and return the final one.
+
+ :param sentence: The input sentence for evalutation
+ :type sentence: basestring
+ :param log: Whether return the score in log representation.
+ :type log: bool
+ :return: Evaluation score, in the decimal or log.
+ :rtype: float
+ """
+ lm = self._language_model_score(sentence)
+ word_cnt = self._word_count(sentence)
+ if log == False:
+ score = np.power(lm, self._alpha) * np.power(word_cnt, self._beta)
+ else:
+ score = self._alpha * np.log(lm) + self._beta * np.log(word_cnt)
+ return score
diff --git a/decoders/swig/__init__.py b/decoders/swig/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/decoders/swig/_init_paths.py b/decoders/swig/_init_paths.py
new file mode 100644
index 000000000..ddabb535b
--- /dev/null
+++ b/decoders/swig/_init_paths.py
@@ -0,0 +1,19 @@
+"""Set up paths for DS2"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os.path
+import sys
+
+
+def add_path(path):
+ if path not in sys.path:
+ sys.path.insert(0, path)
+
+
+this_dir = os.path.dirname(__file__)
+
+# Add project path to PYTHONPATH
+proj_path = os.path.join(this_dir, '..')
+add_path(proj_path)
diff --git a/decoders/swig/ctc_beam_search_decoder.cpp b/decoders/swig/ctc_beam_search_decoder.cpp
new file mode 100644
index 000000000..4a63af26a
--- /dev/null
+++ b/decoders/swig/ctc_beam_search_decoder.cpp
@@ -0,0 +1,222 @@
+#include "ctc_beam_search_decoder.h"
+
+#include
+#include
+#include
+#include
+#include