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, with PaddlePaddle 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 training, and demo deployment. Besides, several pre-trained models for both English and Mandarin are also released.

Table of Contents

• Prerequisites
• Installation
• Getting Started
• Data Preparation
• Training a Model
• Data Augmentation Pipeline
• Inference and Evaluation
• Distributed Cloud Training
• Hyper-parameters Tuning
• Training for Mandarin Language
• Trying Live Demo with Your Own Voice
• Released Models
• Experiments and Benchmarks
• Questions and Help

Prerequisites

• Python 2.7 only supported
• PaddlePaddle the latest version (please refer to the Installation Guide)

Installation

Please make sure the above prerequisites have been satisfied before moving on.

git clone https://github.com/PaddlePaddle/models.git
cd models/deep_speech_2
sh setup.sh

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, Aishell). 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 for instance.

• Go to directory

  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

  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

  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

  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:

  sh run_infer_golden.sh
  

• Evaluate an existing model

  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:

  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 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:

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.

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:

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:

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 a augmentation configuration file in JSON 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:

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 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 and you can download it from statmt. We use part en.00 to train our English languge 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 an internal corpus to train the released Mandarin language model. This corpus contains 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 model only contains Chinese simplified characters. After preprocessing done we can begin to train the language model. The key training arguments are '-o 5 --prune 0 1 2 4 4'. 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:

  CUDA_VISIBLE_DEVICES=0 python infer.py --trainer_count 1
  

• Inference with CPUs:

  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 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:

  CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python test.py --trainer_count 8
  

• Evaluation with CPUs:

  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:

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 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:

  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:

  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

python tune.py --help

or refer to example/librispeech/run_tune.sh.

Distributed Cloud Training

We also provide a cloud training module for users to do the distributed cluster training on PaddleCloud, 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.

Please take the following steps to submit a training job:

• Go to directory:

  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:

  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.

• 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.

• Submit the job:

  By running:

  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:

  paddlecloud get jobs
  

  Run this, the corresponding job's logs will be printed.

  paddlecloud logs -n 10000 $REPLACED_WITH_YOUR_ACTUAL_JOB_NAME
  

For more information about the usage of PaddleCloud, please refer to PaddleCloud Usage.

For more information about the DeepSpeech2 training on PaddleCloud, please refer to Train DeepSpeech2 on PaddleCloud.

Training for Mandarin Language

TODO: to be added

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:

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:

brew install portaudio
pip install pyaudio
pip install pynput

Then to start the client, please run this in another console:

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:

python deploy/demo_server.py --help
python deploy/demo_client.py --help

Released Models

Speech Model Released

Language	Model Name	Training Data	Training Hours
English	LibriSpeech Model	LibriSpeech Dataset	960 h
English	Internal English Model	Baidu English Dataset	8628 h
Mandarin	Aishell Model	Aishell Dataset	151 h
Mandarin	Internal Mandarin Model	Baidu Mandarin Dataset	2917 h

Language Model Released

Language Model	Training Data	Token-based	Size	Filter Configuraiton
English LM	To Be Added	Word-based	8.3 GB	To Be Added
Mandarin LM	To Be Added	Character-based	2.8 GB	To Be Added

Experiments and Benchmarks

English Model Evaluation (Word Error Rate)

Test Set	LibriSpeech Model	Internal English Model
LibriSpeech-Test-Clean	7.96	X.X
LibriSpeech-Test-Other	23.87	X.X
VoxForge-Test	X.X	X.X
Baidu-English-Test	X.X	X.X

(Beam size=2000)

Mandarin Model Evaluation (Character Error Rate)

Test Set	Aishell Model	Internal Mandarin Model
Aishell-Test	X.X	X.X
Baidu-Mandarin-Test	X.X	X.X

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. You are also welcome to contribute to this project.