merge develop to vox12, test=doc

pull/1523/head
xiongxinlei 3 years ago
commit 4473405f82

5
.gitignore vendored

@ -14,6 +14,7 @@
*.whl
*.egg-info
build
*output/
docs/build/
docs/topic/ctc/warp-ctc/
@ -33,6 +34,4 @@ tools/activate_python.sh
tools/miniconda.sh
tools/CRF++-0.58/
speechx/fc_patch/
*output/
speechx/fc_patch/

@ -10,21 +10,15 @@ This demo is an implementation of starting the voice service and accessing the s
### 1. Installation
see [installation](https://github.com/PaddlePaddle/PaddleSpeech/blob/develop/docs/source/install.md).
You can choose one way from easy, meduim and hard to install paddlespeech.
It is recommended to use **paddlepaddle 2.2.1** or above.
You can choose one way from meduim and hard to install paddlespeech.
### 2. Prepare config File
The configuration file contains the service-related configuration files and the model configuration related to the voice tasks contained in the service. They are all under the `conf` folder.
The configuration file can be found in `conf/application.yaml` .
Among them, `engine_list` indicates the speech engine that will be included in the service to be started, in the format of <speech task>_<engine type>.
At present, the speech tasks integrated by the service include: asr (speech recognition) and tts (speech synthesis).
Currently the engine type supports two forms: python and inference (Paddle Inference)
**Note: The configuration of `engine_backend` in `application.yaml` represents all speech tasks included in the started service. **
If the service you want to start contains only a certain speech task, then you need to comment out the speech tasks that do not need to be included. For example, if you only want to use the speech recognition (ASR) service, then you can comment out the speech synthesis (TTS) service, as in the following example:
```bash
engine_backend:
asr: 'conf/asr/asr.yaml'
#tts: 'conf/tts/tts.yaml'
```
**Note: The configuration file of `engine_backend` in `application.yaml` needs to match the configuration type of `engine_type`. **
When the configuration file of `engine_backend` is `XXX.yaml`, the configuration type of `engine_type` needs to be set to `python`; when the configuration file of `engine_backend` is `XXX_pd.yaml`, the configuration of `engine_type` needs to be set type is `inference`;
The input of ASR client demo should be a WAV file(`.wav`), and the sample rate must be the same as the model.

@ -10,19 +10,16 @@
### 1. 安装
请看 [安装文档](https://github.com/PaddlePaddle/PaddleSpeech/blob/develop/docs/source/install.md).
你可以从 easymediumhard 三中方式中选择一种方式安装 PaddleSpeech。
推荐使用 **paddlepaddle 2.2.1** 或以上版本。
你可以从 mediumhard 三中方式中选择一种方式安装 PaddleSpeech。
### 2. 准备配置文件
配置文件包含服务相关的配置文件和服务中包含的语音任务相关的模型配置。 它们都在 `conf` 文件夹下。
**注意:`application.yaml` 中 `engine_backend` 的配置表示启动的服务中包含的所有语音任务。**
如果你想启动的服务中只包含某项语音任务那么你需要注释掉不需要包含的语音任务。例如你只想使用语音识别ASR服务那么你可以将语音合成TTS服务注释掉如下示例
```bash
engine_backend:
asr: 'conf/asr/asr.yaml'
#tts: 'conf/tts/tts.yaml'
```
**注意:`application.yaml` 中 `engine_backend` 的配置文件需要和 `engine_type` 的配置类型匹配。**
当`engine_backend` 的配置文件为`XXX.yaml`时,需要设置`engine_type`的配置类型为`python`;当`engine_backend` 的配置文件为`XXX_pd.yaml`时,需要设置`engine_type`的配置类型为`inference`;
配置文件可参见 `conf/application.yaml`
其中,`engine_list`表示即将启动的服务将会包含的语音引擎,格式为 <语音任务>_<引擎类型>。
目前服务集成的语音任务有: asr(语音识别)、tts(语音合成)。
目前引擎类型支持两种形式python 及 inference (Paddle Inference)
这个 ASR client 的输入应该是一个 WAV 文件(`.wav`),并且采样率必须与模型的采样率相同。
@ -84,7 +81,7 @@ wget -c https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav https://paddlespee
```
### 4. ASR客户端使用方法
**注意:**初次使用客户端时响应时间会略长
**注意:** 初次使用客户端时响应时间会略长
- 命令行 (推荐使用)
```
paddlespeech_client asr --server_ip 127.0.0.1 --port 8090 --input ./zh.wav
@ -133,10 +130,12 @@ wget -c https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav https://paddlespee
```
### 5. TTS客户端使用方法
**注意:**初次使用客户端时响应时间会略长
```bash
paddlespeech_client tts --server_ip 127.0.0.1 --port 8090 --input "您好,欢迎使用百度飞桨语音合成服务。" --output output.wav
```
**注意:** 初次使用客户端时响应时间会略长
- 命令行 (推荐使用)
```bash
paddlespeech_client tts --server_ip 127.0.0.1 --port 8090 --input "您好,欢迎使用百度飞桨语音合成服务。" --output output.wav
```
使用帮助:
```bash

@ -1,27 +1,107 @@
# This is the parameter configuration file for PaddleSpeech Serving.
##################################################################
# SERVER SETTING #
##################################################################
host: '127.0.0.1'
#################################################################################
# SERVER SETTING #
#################################################################################
host: 127.0.0.1
port: 8090
##################################################################
# CONFIG FILE #
##################################################################
# add engine backend type (Options: asr, tts) and config file here.
# Adding a speech task to engine_backend means starting the service.
engine_backend:
asr: 'conf/asr/asr.yaml'
tts: 'conf/tts/tts.yaml'
# The engine_type of speech task needs to keep the same type as the config file of speech task.
# E.g: The engine_type of asr is 'python', the engine_backend of asr is 'XX/asr.yaml'
# E.g: The engine_type of asr is 'inference', the engine_backend of asr is 'XX/asr_pd.yaml'
#
# add engine type (Options: python, inference)
engine_type:
asr: 'python'
tts: 'python'
# The task format in the engin_list is: <speech task>_<engine type>
# task choices = ['asr_python', 'asr_inference', 'tts_python', 'tts_inference']
engine_list: ['asr_python', 'tts_python']
#################################################################################
# ENGINE CONFIG #
#################################################################################
################### speech task: asr; engine_type: python #######################
asr_python:
model: 'conformer_wenetspeech'
lang: 'zh'
sample_rate: 16000
cfg_path: # [optional]
ckpt_path: # [optional]
decode_method: 'attention_rescoring'
force_yes: True
device: # set 'gpu:id' or 'cpu'
################### speech task: asr; engine_type: inference #######################
asr_inference:
# model_type choices=['deepspeech2offline_aishell']
model_type: 'deepspeech2offline_aishell'
am_model: # the pdmodel file of am static model [optional]
am_params: # the pdiparams file of am static model [optional]
lang: 'zh'
sample_rate: 16000
cfg_path:
decode_method:
force_yes: True
am_predictor_conf:
device: # set 'gpu:id' or 'cpu'
switch_ir_optim: True
glog_info: False # True -> print glog
summary: True # False -> do not show predictor config
################### speech task: tts; engine_type: python #######################
tts_python:
# am (acoustic model) choices=['speedyspeech_csmsc', 'fastspeech2_csmsc',
# 'fastspeech2_ljspeech', 'fastspeech2_aishell3',
# 'fastspeech2_vctk']
am: 'fastspeech2_csmsc'
am_config:
am_ckpt:
am_stat:
phones_dict:
tones_dict:
speaker_dict:
spk_id: 0
# voc (vocoder) choices=['pwgan_csmsc', 'pwgan_ljspeech', 'pwgan_aishell3',
# 'pwgan_vctk', 'mb_melgan_csmsc']
voc: 'pwgan_csmsc'
voc_config:
voc_ckpt:
voc_stat:
# others
lang: 'zh'
device: # set 'gpu:id' or 'cpu'
################### speech task: tts; engine_type: inference #######################
tts_inference:
# am (acoustic model) choices=['speedyspeech_csmsc', 'fastspeech2_csmsc']
am: 'fastspeech2_csmsc'
am_model: # the pdmodel file of your am static model (XX.pdmodel)
am_params: # the pdiparams file of your am static model (XX.pdipparams)
am_sample_rate: 24000
phones_dict:
tones_dict:
speaker_dict:
spk_id: 0
am_predictor_conf:
device: # set 'gpu:id' or 'cpu'
switch_ir_optim: True
glog_info: False # True -> print glog
summary: True # False -> do not show predictor config
# voc (vocoder) choices=['pwgan_csmsc', 'mb_melgan_csmsc','hifigan_csmsc']
voc: 'pwgan_csmsc'
voc_model: # the pdmodel file of your vocoder static model (XX.pdmodel)
voc_params: # the pdiparams file of your vocoder static model (XX.pdipparams)
voc_sample_rate: 24000
voc_predictor_conf:
device: # set 'gpu:id' or 'cpu'
switch_ir_optim: True
glog_info: False # True -> print glog
summary: True # False -> do not show predictor config
# others
lang: 'zh'

@ -1,8 +0,0 @@
model: 'conformer_wenetspeech'
lang: 'zh'
sample_rate: 16000
cfg_path: # [optional]
ckpt_path: # [optional]
decode_method: 'attention_rescoring'
force_yes: True
device: # set 'gpu:id' or 'cpu'

@ -1,26 +0,0 @@
# This is the parameter configuration file for ASR server.
# These are the static models that support paddle inference.
##################################################################
# ACOUSTIC MODEL SETTING #
# am choices=['deepspeech2offline_aishell'] TODO
##################################################################
model_type: 'deepspeech2offline_aishell'
am_model: # the pdmodel file of am static model [optional]
am_params: # the pdiparams file of am static model [optional]
lang: 'zh'
sample_rate: 16000
cfg_path:
decode_method:
force_yes: True
am_predictor_conf:
device: # set 'gpu:id' or 'cpu'
switch_ir_optim: True
glog_info: False # True -> print glog
summary: True # False -> do not show predictor config
##################################################################
# OTHERS #
##################################################################

@ -1,32 +0,0 @@
# This is the parameter configuration file for TTS server.
##################################################################
# ACOUSTIC MODEL SETTING #
# am choices=['speedyspeech_csmsc', 'fastspeech2_csmsc',
# 'fastspeech2_ljspeech', 'fastspeech2_aishell3',
# 'fastspeech2_vctk']
##################################################################
am: 'fastspeech2_csmsc'
am_config:
am_ckpt:
am_stat:
phones_dict:
tones_dict:
speaker_dict:
spk_id: 0
##################################################################
# VOCODER SETTING #
# voc choices=['pwgan_csmsc', 'pwgan_ljspeech', 'pwgan_aishell3',
# 'pwgan_vctk', 'mb_melgan_csmsc']
##################################################################
voc: 'pwgan_csmsc'
voc_config:
voc_ckpt:
voc_stat:
##################################################################
# OTHERS #
##################################################################
lang: 'zh'
device: # set 'gpu:id' or 'cpu'

@ -1,42 +0,0 @@
# This is the parameter configuration file for TTS server.
# These are the static models that support paddle inference.
##################################################################
# ACOUSTIC MODEL SETTING #
# am choices=['speedyspeech_csmsc', 'fastspeech2_csmsc']
##################################################################
am: 'fastspeech2_csmsc'
am_model: # the pdmodel file of your am static model (XX.pdmodel)
am_params: # the pdiparams file of your am static model (XX.pdipparams)
am_sample_rate: 24000
phones_dict:
tones_dict:
speaker_dict:
spk_id: 0
am_predictor_conf:
device: # set 'gpu:id' or 'cpu'
switch_ir_optim: True
glog_info: False # True -> print glog
summary: True # False -> do not show predictor config
##################################################################
# VOCODER SETTING #
# voc choices=['pwgan_csmsc', 'mb_melgan_csmsc','hifigan_csmsc']
##################################################################
voc: 'pwgan_csmsc'
voc_model: # the pdmodel file of your vocoder static model (XX.pdmodel)
voc_params: # the pdiparams file of your vocoder static model (XX.pdipparams)
voc_sample_rate: 24000
voc_predictor_conf:
device: # set 'gpu:id' or 'cpu'
switch_ir_optim: True
glog_info: False # True -> print glog
summary: True # False -> do not show predictor config
##################################################################
# OTHERS #
##################################################################
lang: 'zh'

@ -1,3 +1,3 @@
#!/bin/bash
paddlespeech_server start --config_file ./conf/application.yaml
paddlespeech_server start --config_file ./conf/application.yaml

@ -0,0 +1,142 @@
# HiFiGAN with AISHELL-3
This example contains code used to train a [HiFiGAN](https://arxiv.org/abs/2010.05646) model with [AISHELL-3](http://www.aishelltech.com/aishell_3).
AISHELL-3 is a large-scale and high-fidelity multi-speaker Mandarin speech corpus that could be used to train multi-speaker Text-to-Speech (TTS) systems.
## Dataset
### Download and Extract
Download AISHELL-3.
```bash
wget https://www.openslr.org/resources/93/data_aishell3.tgz
```
Extract AISHELL-3.
```bash
mkdir data_aishell3
tar zxvf data_aishell3.tgz -C data_aishell3
```
### Get MFA Result and Extract
We use [MFA2.x](https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner) to get durations for aishell3_fastspeech2.
You can download from here [aishell3_alignment_tone.tar.gz](https://paddlespeech.bj.bcebos.com/MFA/AISHELL-3/with_tone/aishell3_alignment_tone.tar.gz), or train your MFA model reference to [mfa example](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/mfa) (use MFA1.x now) of our repo.
## Get Started
Assume the path to the dataset is `~/datasets/data_aishell3`.
Assume the path to the MFA result of AISHELL-3 is `./aishell3_alignment_tone`.
Run the command below to
1. **source path**.
2. preprocess the dataset.
3. train the model.
4. synthesize wavs.
- synthesize waveform from `metadata.jsonl`.
```bash
./run.sh
```
You can choose a range of stages you want to run, or set `stage` equal to `stop-stage` to use only one stage, for example, run the following command will only preprocess the dataset.
```bash
./run.sh --stage 0 --stop-stage 0
```
### Data Preprocessing
```bash
./local/preprocess.sh ${conf_path}
```
When it is done. A `dump` folder is created in the current directory. The structure of the dump folder is listed below.
```text
dump
├── dev
│ ├── norm
│ └── raw
├── test
│ ├── norm
│ └── raw
└── train
├── norm
├── raw
└── feats_stats.npy
```
The dataset is split into 3 parts, namely `train`, `dev`, and `test`, each of which contains a `norm` and `raw` subfolder. The `raw` folder contains the log magnitude of the mel spectrogram of each utterance, while the norm folder contains the normalized spectrogram. The statistics used to normalize the spectrogram are computed from the training set, which is located in `dump/train/feats_stats.npy`.
Also, there is a `metadata.jsonl` in each subfolder. It is a table-like file that contains id and paths to the spectrogram of each utterance.
### Model Training
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${train_output_path}
```
`./local/train.sh` calls `${BIN_DIR}/train.py`.
Here's the complete help message.
```text
usage: train.py [-h] [--config CONFIG] [--train-metadata TRAIN_METADATA]
[--dev-metadata DEV_METADATA] [--output-dir OUTPUT_DIR]
[--ngpu NGPU] [--batch-size BATCH_SIZE] [--max-iter MAX_ITER]
[--run-benchmark RUN_BENCHMARK]
[--profiler_options PROFILER_OPTIONS]
Train a ParallelWaveGAN model.
optional arguments:
-h, --help show this help message and exit
--config CONFIG config file to overwrite default config.
--train-metadata TRAIN_METADATA
training data.
--dev-metadata DEV_METADATA
dev data.
--output-dir OUTPUT_DIR
output dir.
--ngpu NGPU if ngpu == 0, use cpu.
benchmark:
arguments related to benchmark.
--batch-size BATCH_SIZE
batch size.
--max-iter MAX_ITER train max steps.
--run-benchmark RUN_BENCHMARK
runing benchmark or not, if True, use the --batch-size
and --max-iter.
--profiler_options PROFILER_OPTIONS
The option of profiler, which should be in format
"key1=value1;key2=value2;key3=value3".
```
1. `--config` is a config file in yaml format to overwrite the default config, which can be found at `conf/default.yaml`.
2. `--train-metadata` and `--dev-metadata` should be the metadata file in the normalized subfolder of `train` and `dev` in the `dump` folder.
3. `--output-dir` is the directory to save the results of the experiment. Checkpoints are saved in `checkpoints/` inside this directory.
4. `--ngpu` is the number of gpus to use, if ngpu == 0, use cpu.
### Synthesizing
`./local/synthesize.sh` calls `${BIN_DIR}/../synthesize.py`, which can synthesize waveform from `metadata.jsonl`.
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name}
```
```text
usage: synthesize.py [-h] [--generator-type GENERATOR_TYPE] [--config CONFIG]
[--checkpoint CHECKPOINT] [--test-metadata TEST_METADATA]
[--output-dir OUTPUT_DIR] [--ngpu NGPU]
Synthesize with GANVocoder.
optional arguments:
-h, --help show this help message and exit
--generator-type GENERATOR_TYPE
type of GANVocoder, should in {pwgan, mb_melgan,
style_melgan, } now
--config CONFIG GANVocoder config file.
--checkpoint CHECKPOINT
snapshot to load.
--test-metadata TEST_METADATA
dev data.
--output-dir OUTPUT_DIR
output dir.
--ngpu NGPU if ngpu == 0, use cpu.
```
1. `--config` config file. You should use the same config with which the model is trained.
2. `--checkpoint` is the checkpoint to load. Pick one of the checkpoints from `checkpoints` inside the training output directory.
3. `--test-metadata` is the metadata of the test dataset. Use the `metadata.jsonl` in the `dev/norm` subfolder from the processed directory.
4. `--output-dir` is the directory to save the synthesized audio files.
5. `--ngpu` is the number of gpus to use, if ngpu == 0, use cpu.
## Pretrained Models
## Acknowledgement
We adapted some code from https://github.com/kan-bayashi/ParallelWaveGAN.

@ -0,0 +1,168 @@
# This is the configuration file for AISHELL-3 dataset.
# This configuration is based on HiFiGAN V1, which is
# an official configuration. But I found that the optimizer
# setting does not work well with my implementation.
# So I changed optimizer settings as follows:
# - AdamW -> Adam
# - betas: [0.8, 0.99] -> betas: [0.5, 0.9]
# - Scheduler: ExponentialLR -> MultiStepLR
# To match the shift size difference, the upsample scales
# is also modified from the original 256 shift setting.
###########################################################
# FEATURE EXTRACTION SETTING #
###########################################################
fs: 24000 # Sampling rate.
n_fft: 2048 # FFT size (samples).
n_shift: 300 # Hop size (samples). 12.5ms
win_length: 1200 # Window length (samples). 50ms
# If set to null, it will be the same as fft_size.
window: "hann" # Window function.
n_mels: 80 # Number of mel basis.
fmin: 80 # Minimum freq in mel basis calculation. (Hz)
fmax: 7600 # Maximum frequency in mel basis calculation. (Hz)
###########################################################
# GENERATOR NETWORK ARCHITECTURE SETTING #
###########################################################
generator_params:
in_channels: 80 # Number of input channels.
out_channels: 1 # Number of output channels.
channels: 512 # Number of initial channels.
kernel_size: 7 # Kernel size of initial and final conv layers.
upsample_scales: [5, 5, 4, 3] # Upsampling scales.
upsample_kernel_sizes: [10, 10, 8, 6] # Kernel size for upsampling layers.
resblock_kernel_sizes: [3, 7, 11] # Kernel size for residual blocks.
resblock_dilations: # Dilations for residual blocks.
- [1, 3, 5]
- [1, 3, 5]
- [1, 3, 5]
use_additional_convs: True # Whether to use additional conv layer in residual blocks.
bias: True # Whether to use bias parameter in conv.
nonlinear_activation: "leakyrelu" # Nonlinear activation type.
nonlinear_activation_params: # Nonlinear activation paramters.
negative_slope: 0.1
use_weight_norm: True # Whether to apply weight normalization.
###########################################################
# DISCRIMINATOR NETWORK ARCHITECTURE SETTING #
###########################################################
discriminator_params:
scales: 3 # Number of multi-scale discriminator.
scale_downsample_pooling: "AvgPool1D" # Pooling operation for scale discriminator.
scale_downsample_pooling_params:
kernel_size: 4 # Pooling kernel size.
stride: 2 # Pooling stride.
padding: 2 # Padding size.
scale_discriminator_params:
in_channels: 1 # Number of input channels.
out_channels: 1 # Number of output channels.
kernel_sizes: [15, 41, 5, 3] # List of kernel sizes.
channels: 128 # Initial number of channels.
max_downsample_channels: 1024 # Maximum number of channels in downsampling conv layers.
max_groups: 16 # Maximum number of groups in downsampling conv layers.
bias: True
downsample_scales: [4, 4, 4, 4, 1] # Downsampling scales.
nonlinear_activation: "leakyrelu" # Nonlinear activation.
nonlinear_activation_params:
negative_slope: 0.1
follow_official_norm: True # Whether to follow the official norm setting.
periods: [2, 3, 5, 7, 11] # List of period for multi-period discriminator.
period_discriminator_params:
in_channels: 1 # Number of input channels.
out_channels: 1 # Number of output channels.
kernel_sizes: [5, 3] # List of kernel sizes.
channels: 32 # Initial number of channels.
downsample_scales: [3, 3, 3, 3, 1] # Downsampling scales.
max_downsample_channels: 1024 # Maximum number of channels in downsampling conv layers.
bias: True # Whether to use bias parameter in conv layer."
nonlinear_activation: "leakyrelu" # Nonlinear activation.
nonlinear_activation_params: # Nonlinear activation paramters.
negative_slope: 0.1
use_weight_norm: True # Whether to apply weight normalization.
use_spectral_norm: False # Whether to apply spectral normalization.
###########################################################
# STFT LOSS SETTING #
###########################################################
use_stft_loss: False # Whether to use multi-resolution STFT loss.
use_mel_loss: True # Whether to use Mel-spectrogram loss.
mel_loss_params:
fs: 24000
fft_size: 2048
hop_size: 300
win_length: 1200
window: "hann"
num_mels: 80
fmin: 0
fmax: 12000
log_base: null
generator_adv_loss_params:
average_by_discriminators: False # Whether to average loss by #discriminators.
discriminator_adv_loss_params:
average_by_discriminators: False # Whether to average loss by #discriminators.
use_feat_match_loss: True
feat_match_loss_params:
average_by_discriminators: False # Whether to average loss by #discriminators.
average_by_layers: False # Whether to average loss by #layers in each discriminator.
include_final_outputs: False # Whether to include final outputs in feat match loss calculation.
###########################################################
# ADVERSARIAL LOSS SETTING #
###########################################################
lambda_aux: 45.0 # Loss balancing coefficient for STFT loss.
lambda_adv: 1.0 # Loss balancing coefficient for adversarial loss.
lambda_feat_match: 2.0 # Loss balancing coefficient for feat match loss..
###########################################################
# DATA LOADER SETTING #
###########################################################
batch_size: 16 # Batch size.
batch_max_steps: 8400 # Length of each audio in batch. Make sure dividable by hop_size.
num_workers: 2 # Number of workers in DataLoader.
###########################################################
# OPTIMIZER & SCHEDULER SETTING #
###########################################################
generator_optimizer_params:
beta1: 0.5
beta2: 0.9
weight_decay: 0.0 # Generator's weight decay coefficient.
generator_scheduler_params:
learning_rate: 2.0e-4 # Generator's learning rate.
gamma: 0.5 # Generator's scheduler gamma.
milestones: # At each milestone, lr will be multiplied by gamma.
- 200000
- 400000
- 600000
- 800000
generator_grad_norm: -1 # Generator's gradient norm.
discriminator_optimizer_params:
beta1: 0.5
beta2: 0.9
weight_decay: 0.0 # Discriminator's weight decay coefficient.
discriminator_scheduler_params:
learning_rate: 2.0e-4 # Discriminator's learning rate.
gamma: 0.5 # Discriminator's scheduler gamma.
milestones: # At each milestone, lr will be multiplied by gamma.
- 200000
- 400000
- 600000
- 800000
discriminator_grad_norm: -1 # Discriminator's gradient norm.
###########################################################
# INTERVAL SETTING #
###########################################################
generator_train_start_steps: 1 # Number of steps to start to train discriminator.
discriminator_train_start_steps: 0 # Number of steps to start to train discriminator.
train_max_steps: 2500000 # Number of training steps.
save_interval_steps: 5000 # Interval steps to save checkpoint.
eval_interval_steps: 1000 # Interval steps to evaluate the network.
###########################################################
# OTHER SETTING #
###########################################################
num_snapshots: 10 # max number of snapshots to keep while training
seed: 42 # random seed for paddle, random, and np.random

@ -0,0 +1,55 @@
#!/bin/bash
stage=0
stop_stage=100
config_path=$1
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
# get durations from MFA's result
echo "Generate durations.txt from MFA results ..."
python3 ${MAIN_ROOT}/utils/gen_duration_from_textgrid.py \
--inputdir=./aishell3_alignment_tone \
--output=durations.txt \
--config=${config_path}
fi
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
# extract features
echo "Extract features ..."
python3 ${BIN_DIR}/../preprocess.py \
--rootdir=~/datasets/data_aishell3/ \
--dataset=aishell3 \
--dumpdir=dump \
--dur-file=durations.txt \
--config=${config_path} \
--cut-sil=True \
--num-cpu=20
fi
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
# get features' stats(mean and std)
echo "Get features' stats ..."
python3 ${MAIN_ROOT}/utils/compute_statistics.py \
--metadata=dump/train/raw/metadata.jsonl \
--field-name="feats"
fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
# normalize, dev and test should use train's stats
echo "Normalize ..."
python3 ${BIN_DIR}/../normalize.py \
--metadata=dump/train/raw/metadata.jsonl \
--dumpdir=dump/train/norm \
--stats=dump/train/feats_stats.npy
python3 ${BIN_DIR}/../normalize.py \
--metadata=dump/dev/raw/metadata.jsonl \
--dumpdir=dump/dev/norm \
--stats=dump/train/feats_stats.npy
python3 ${BIN_DIR}/../normalize.py \
--metadata=dump/test/raw/metadata.jsonl \
--dumpdir=dump/test/norm \
--stats=dump/train/feats_stats.npy
fi

@ -0,0 +1,14 @@
#!/bin/bash
config_path=$1
train_output_path=$2
ckpt_name=$3
FLAGS_allocator_strategy=naive_best_fit \
FLAGS_fraction_of_gpu_memory_to_use=0.01 \
python3 ${BIN_DIR}/../synthesize.py \
--config=${config_path} \
--checkpoint=${train_output_path}/checkpoints/${ckpt_name} \
--test-metadata=dump/test/norm/metadata.jsonl \
--output-dir=${train_output_path}/test \
--generator-type=hifigan

@ -0,0 +1,13 @@
#!/bin/bash
config_path=$1
train_output_path=$2
FLAGS_cudnn_exhaustive_search=true \
FLAGS_conv_workspace_size_limit=4000 \
python ${BIN_DIR}/train.py \
--train-metadata=dump/train/norm/metadata.jsonl \
--dev-metadata=dump/dev/norm/metadata.jsonl \
--config=${config_path} \
--output-dir=${train_output_path} \
--ngpu=1

@ -0,0 +1,13 @@
#!/bin/bash
export MAIN_ROOT=`realpath ${PWD}/../../../`
export PATH=${MAIN_ROOT}:${MAIN_ROOT}/utils:${PATH}
export LC_ALL=C
export PYTHONDONTWRITEBYTECODE=1
# Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
export PYTHONIOENCODING=UTF-8
export PYTHONPATH=${MAIN_ROOT}:${PYTHONPATH}
MODEL=hifigan
export BIN_DIR=${MAIN_ROOT}/paddlespeech/t2s/exps/gan_vocoder/${MODEL}

@ -0,0 +1,32 @@
#!/bin/bash
set -e
source path.sh
gpus=0
stage=0
stop_stage=100
conf_path=conf/default.yaml
train_output_path=exp/default
ckpt_name=snapshot_iter_5000.pdz
# with the following command, you can choose the stage range you want to run
# such as `./run.sh --stage 0 --stop-stage 0`
# this can not be mixed use with `$1`, `$2` ...
source ${MAIN_ROOT}/utils/parse_options.sh || exit 1
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
# prepare data
./local/preprocess.sh ${conf_path} || exit -1
fi
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
# train model, all `ckpt` under `train_output_path/checkpoints/` dir
CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${train_output_path} || exit -1
fi
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
# synthesize
CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name} || exit -1
fi

@ -0,0 +1,139 @@
# HiFiGAN with VCTK
This example contains code used to train a [HiFiGAN](https://arxiv.org/abs/2010.05646) model with [VCTK](https://datashare.ed.ac.uk/handle/10283/3443).
## Dataset
### Download and Extract
Download VCTK-0.92 from the [official website](https://datashare.ed.ac.uk/handle/10283/3443) and extract it to `~/datasets`. Then the dataset is in directory `~/datasets/VCTK-Corpus-0.92`.
### Get MFA Result and Extract
We use [MFA](https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner) results to cut the silence in the edge of audio.
You can download from here [vctk_alignment.tar.gz](https://paddlespeech.bj.bcebos.com/MFA/VCTK-Corpus-0.92/vctk_alignment.tar.gz), or train your MFA model reference to [mfa example](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/examples/other/mfa) of our repo.
ps: we remove three speakers in VCTK-0.92 (see [reorganize_vctk.py](https://github.com/PaddlePaddle/PaddleSpeech/blob/develop/examples/other/mfa/local/reorganize_vctk.py)):
1. `p315`, because of no text for it.
2. `p280` and `p362`, because no *_mic2.flac (which is better than *_mic1.flac) for them.
## Get Started
Assume the path to the dataset is `~/datasets/VCTK-Corpus-0.92`.
Assume the path to the MFA result of VCTK is `./vctk_alignment`.
Run the command below to
1. **source path**.
2. preprocess the dataset.
3. train the model.
4. synthesize wavs.
- synthesize waveform from `metadata.jsonl`.
```bash
./run.sh
```
You can choose a range of stages you want to run, or set `stage` equal to `stop-stage` to use only one stage, for example, running the following command will only preprocess the dataset.
```bash
./run.sh --stage 0 --stop-stage 0
```
### Data Preprocessing
```bash
./local/preprocess.sh ${conf_path}
```
When it is done. A `dump` folder is created in the current directory. The structure of the dump folder is listed below.
```text
dump
├── dev
│ ├── norm
│ └── raw
├── test
│ ├── norm
│ └── raw
└── train
├── norm
├── raw
└── feats_stats.npy
```
The dataset is split into 3 parts, namely `train`, `dev`, and `test`, each of which contains a `norm` and `raw` subfolder. The `raw` folder contains the log magnitude of the mel spectrogram of each utterance, while the norm folder contains the normalized spectrogram. The statistics used to normalize the spectrogram are computed from the training set, which is located in `dump/train/feats_stats.npy`.
Also, there is a `metadata.jsonl` in each subfolder. It is a table-like file that contains id and paths to the spectrogram of each utterance.
### Model Training
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${train_output_path}
```
`./local/train.sh` calls `${BIN_DIR}/train.py`.
Here's the complete help message.
```text
usage: train.py [-h] [--config CONFIG] [--train-metadata TRAIN_METADATA]
[--dev-metadata DEV_METADATA] [--output-dir OUTPUT_DIR]
[--ngpu NGPU] [--batch-size BATCH_SIZE] [--max-iter MAX_ITER]
[--run-benchmark RUN_BENCHMARK]
[--profiler_options PROFILER_OPTIONS]
Train a ParallelWaveGAN model.
optional arguments:
-h, --help show this help message and exit
--config CONFIG config file to overwrite default config.
--train-metadata TRAIN_METADATA
training data.
--dev-metadata DEV_METADATA
dev data.
--output-dir OUTPUT_DIR
output dir.
--ngpu NGPU if ngpu == 0, use cpu.
benchmark:
arguments related to benchmark.
--batch-size BATCH_SIZE
batch size.
--max-iter MAX_ITER train max steps.
--run-benchmark RUN_BENCHMARK
runing benchmark or not, if True, use the --batch-size
and --max-iter.
--profiler_options PROFILER_OPTIONS
The option of profiler, which should be in format
"key1=value1;key2=value2;key3=value3".
```
1. `--config` is a config file in yaml format to overwrite the default config, which can be found at `conf/default.yaml`.
2. `--train-metadata` and `--dev-metadata` should be the metadata file in the normalized subfolder of `train` and `dev` in the `dump` folder.
3. `--output-dir` is the directory to save the results of the experiment. Checkpoints are saved in `checkpoints/` inside this directory.
4. `--ngpu` is the number of gpus to use, if ngpu == 0, use cpu.
### Synthesizing
`./local/synthesize.sh` calls `${BIN_DIR}/../synthesize.py`, which can synthesize waveform from `metadata.jsonl`.
```bash
CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name}
```
```text
usage: synthesize.py [-h] [--generator-type GENERATOR_TYPE] [--config CONFIG]
[--checkpoint CHECKPOINT] [--test-metadata TEST_METADATA]
[--output-dir OUTPUT_DIR] [--ngpu NGPU]
Synthesize with GANVocoder.
optional arguments:
-h, --help show this help message and exit
--generator-type GENERATOR_TYPE
type of GANVocoder, should in {pwgan, mb_melgan,
style_melgan, } now
--config CONFIG GANVocoder config file.
--checkpoint CHECKPOINT
snapshot to load.
--test-metadata TEST_METADATA
dev data.
--output-dir OUTPUT_DIR
output dir.
--ngpu NGPU if ngpu == 0, use cpu.
```
1. `--config` config file. You should use the same config with which the model is trained.
2. `--checkpoint` is the checkpoint to load. Pick one of the checkpoints from `checkpoints` inside the training output directory.
3. `--test-metadata` is the metadata of the test dataset. Use the `metadata.jsonl` in the `dev/norm` subfolder from the processed directory.
4. `--output-dir` is the directory to save the synthesized audio files.
5. `--ngpu` is the number of gpus to use, if ngpu == 0, use cpu.
## Pretrained Model
## Acknowledgement
We adapted some code from https://github.com/kan-bayashi/ParallelWaveGAN.

@ -0,0 +1,168 @@
# This is the configuration file for VCTK dataset.
# This configuration is based on HiFiGAN V1, which is
# an official configuration. But I found that the optimizer
# setting does not work well with my implementation.
# So I changed optimizer settings as follows:
# - AdamW -> Adam
# - betas: [0.8, 0.99] -> betas: [0.5, 0.9]
# - Scheduler: ExponentialLR -> MultiStepLR
# To match the shift size difference, the upsample scales
# is also modified from the original 256 shift setting.
###########################################################
# FEATURE EXTRACTION SETTING #
###########################################################
fs: 24000 # Sampling rate.
n_fft: 2048 # FFT size (samples).
n_shift: 300 # Hop size (samples). 12.5ms
win_length: 1200 # Window length (samples). 50ms
# If set to null, it will be the same as fft_size.
window: "hann" # Window function.
n_mels: 80 # Number of mel basis.
fmin: 80 # Minimum freq in mel basis calculation. (Hz)
fmax: 7600 # Maximum frequency in mel basis calculation. (Hz)
###########################################################
# GENERATOR NETWORK ARCHITECTURE SETTING #
###########################################################
generator_params:
in_channels: 80 # Number of input channels.
out_channels: 1 # Number of output channels.
channels: 512 # Number of initial channels.
kernel_size: 7 # Kernel size of initial and final conv layers.
upsample_scales: [5, 5, 4, 3] # Upsampling scales.
upsample_kernel_sizes: [10, 10, 8, 6] # Kernel size for upsampling layers.
resblock_kernel_sizes: [3, 7, 11] # Kernel size for residual blocks.
resblock_dilations: # Dilations for residual blocks.
- [1, 3, 5]
- [1, 3, 5]
- [1, 3, 5]
use_additional_convs: True # Whether to use additional conv layer in residual blocks.
bias: True # Whether to use bias parameter in conv.
nonlinear_activation: "leakyrelu" # Nonlinear activation type.
nonlinear_activation_params: # Nonlinear activation paramters.
negative_slope: 0.1
use_weight_norm: True # Whether to apply weight normalization.
###########################################################
# DISCRIMINATOR NETWORK ARCHITECTURE SETTING #
###########################################################
discriminator_params:
scales: 3 # Number of multi-scale discriminator.
scale_downsample_pooling: "AvgPool1D" # Pooling operation for scale discriminator.
scale_downsample_pooling_params:
kernel_size: 4 # Pooling kernel size.
stride: 2 # Pooling stride.
padding: 2 # Padding size.
scale_discriminator_params:
in_channels: 1 # Number of input channels.
out_channels: 1 # Number of output channels.
kernel_sizes: [15, 41, 5, 3] # List of kernel sizes.
channels: 128 # Initial number of channels.
max_downsample_channels: 1024 # Maximum number of channels in downsampling conv layers.
max_groups: 16 # Maximum number of groups in downsampling conv layers.
bias: True
downsample_scales: [4, 4, 4, 4, 1] # Downsampling scales.
nonlinear_activation: "leakyrelu" # Nonlinear activation.
nonlinear_activation_params:
negative_slope: 0.1
follow_official_norm: True # Whether to follow the official norm setting.
periods: [2, 3, 5, 7, 11] # List of period for multi-period discriminator.
period_discriminator_params:
in_channels: 1 # Number of input channels.
out_channels: 1 # Number of output channels.
kernel_sizes: [5, 3] # List of kernel sizes.
channels: 32 # Initial number of channels.
downsample_scales: [3, 3, 3, 3, 1] # Downsampling scales.
max_downsample_channels: 1024 # Maximum number of channels in downsampling conv layers.
bias: True # Whether to use bias parameter in conv layer."
nonlinear_activation: "leakyrelu" # Nonlinear activation.
nonlinear_activation_params: # Nonlinear activation paramters.
negative_slope: 0.1
use_weight_norm: True # Whether to apply weight normalization.
use_spectral_norm: False # Whether to apply spectral normalization.
###########################################################
# STFT LOSS SETTING #
###########################################################
use_stft_loss: False # Whether to use multi-resolution STFT loss.
use_mel_loss: True # Whether to use Mel-spectrogram loss.
mel_loss_params:
fs: 24000
fft_size: 2048
hop_size: 300
win_length: 1200
window: "hann"
num_mels: 80
fmin: 0
fmax: 12000
log_base: null
generator_adv_loss_params:
average_by_discriminators: False # Whether to average loss by #discriminators.
discriminator_adv_loss_params:
average_by_discriminators: False # Whether to average loss by #discriminators.
use_feat_match_loss: True
feat_match_loss_params:
average_by_discriminators: False # Whether to average loss by #discriminators.
average_by_layers: False # Whether to average loss by #layers in each discriminator.
include_final_outputs: False # Whether to include final outputs in feat match loss calculation.
###########################################################
# ADVERSARIAL LOSS SETTING #
###########################################################
lambda_aux: 45.0 # Loss balancing coefficient for STFT loss.
lambda_adv: 1.0 # Loss balancing coefficient for adversarial loss.
lambda_feat_match: 2.0 # Loss balancing coefficient for feat match loss..
###########################################################
# DATA LOADER SETTING #
###########################################################
batch_size: 16 # Batch size.
batch_max_steps: 8400 # Length of each audio in batch. Make sure dividable by hop_size.
num_workers: 2 # Number of workers in DataLoader.
###########################################################
# OPTIMIZER & SCHEDULER SETTING #
###########################################################
generator_optimizer_params:
beta1: 0.5
beta2: 0.9
weight_decay: 0.0 # Generator's weight decay coefficient.
generator_scheduler_params:
learning_rate: 2.0e-4 # Generator's learning rate.
gamma: 0.5 # Generator's scheduler gamma.
milestones: # At each milestone, lr will be multiplied by gamma.
- 200000
- 400000
- 600000
- 800000
generator_grad_norm: -1 # Generator's gradient norm.
discriminator_optimizer_params:
beta1: 0.5
beta2: 0.9
weight_decay: 0.0 # Discriminator's weight decay coefficient.
discriminator_scheduler_params:
learning_rate: 2.0e-4 # Discriminator's learning rate.
gamma: 0.5 # Discriminator's scheduler gamma.
milestones: # At each milestone, lr will be multiplied by gamma.
- 200000
- 400000
- 600000
- 800000
discriminator_grad_norm: -1 # Discriminator's gradient norm.
###########################################################
# INTERVAL SETTING #
###########################################################
generator_train_start_steps: 1 # Number of steps to start to train discriminator.
discriminator_train_start_steps: 0 # Number of steps to start to train discriminator.
train_max_steps: 2500000 # Number of training steps.
save_interval_steps: 5000 # Interval steps to save checkpoint.
eval_interval_steps: 1000 # Interval steps to evaluate the network.
###########################################################
# OTHER SETTING #
###########################################################
num_snapshots: 10 # max number of snapshots to keep while training
seed: 42 # random seed for paddle, random, and np.random

@ -0,0 +1,55 @@
#!/bin/bash
stage=0
stop_stage=100
config_path=$1
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
# get durations from MFA's result
echo "Generate durations.txt from MFA results ..."
python3 ${MAIN_ROOT}/utils/gen_duration_from_textgrid.py \
--inputdir=./vctk_alignment \
--output=durations.txt \
--config=${config_path}
fi
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
# extract features
echo "Extract features ..."
python3 ${BIN_DIR}/../preprocess.py \
--rootdir=~/datasets/VCTK-Corpus-0.92/ \
--dataset=vctk \
--dumpdir=dump \
--dur-file=durations.txt \
--config=${config_path} \
--cut-sil=True \
--num-cpu=20
fi
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
# get features' stats(mean and std)
echo "Get features' stats ..."
python3 ${MAIN_ROOT}/utils/compute_statistics.py \
--metadata=dump/train/raw/metadata.jsonl \
--field-name="feats"
fi
if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
# normalize, dev and test should use train's stats
echo "Normalize ..."
python3 ${BIN_DIR}/../normalize.py \
--metadata=dump/train/raw/metadata.jsonl \
--dumpdir=dump/train/norm \
--stats=dump/train/feats_stats.npy
python3 ${BIN_DIR}/../normalize.py \
--metadata=dump/dev/raw/metadata.jsonl \
--dumpdir=dump/dev/norm \
--stats=dump/train/feats_stats.npy
python3 ${BIN_DIR}/../normalize.py \
--metadata=dump/test/raw/metadata.jsonl \
--dumpdir=dump/test/norm \
--stats=dump/train/feats_stats.npy
fi

@ -0,0 +1,14 @@
#!/bin/bash
config_path=$1
train_output_path=$2
ckpt_name=$3
FLAGS_allocator_strategy=naive_best_fit \
FLAGS_fraction_of_gpu_memory_to_use=0.01 \
python3 ${BIN_DIR}/../synthesize.py \
--config=${config_path} \
--checkpoint=${train_output_path}/checkpoints/${ckpt_name} \
--test-metadata=dump/test/norm/metadata.jsonl \
--output-dir=${train_output_path}/test \
--generator-type=hifigan

@ -0,0 +1,13 @@
#!/bin/bash
config_path=$1
train_output_path=$2
FLAGS_cudnn_exhaustive_search=true \
FLAGS_conv_workspace_size_limit=4000 \
python ${BIN_DIR}/train.py \
--train-metadata=dump/train/norm/metadata.jsonl \
--dev-metadata=dump/dev/norm/metadata.jsonl \
--config=${config_path} \
--output-dir=${train_output_path} \
--ngpu=1

@ -0,0 +1,13 @@
#!/bin/bash
export MAIN_ROOT=`realpath ${PWD}/../../../`
export PATH=${MAIN_ROOT}:${MAIN_ROOT}/utils:${PATH}
export LC_ALL=C
export PYTHONDONTWRITEBYTECODE=1
# Use UTF-8 in Python to avoid UnicodeDecodeError when LC_ALL=C
export PYTHONIOENCODING=UTF-8
export PYTHONPATH=${MAIN_ROOT}:${PYTHONPATH}
MODEL=hifigan
export BIN_DIR=${MAIN_ROOT}/paddlespeech/t2s/exps/gan_vocoder/${MODEL}

@ -0,0 +1,32 @@
#!/bin/bash
set -e
source path.sh
gpus=0
stage=0
stop_stage=100
conf_path=conf/default.yaml
train_output_path=exp/default
ckpt_name=snapshot_iter_5000.pdz
# with the following command, you can choose the stage range you want to run
# such as `./run.sh --stage 0 --stop-stage 0`
# this can not be mixed use with `$1`, `$2` ...
source ${MAIN_ROOT}/utils/parse_options.sh || exit 1
if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
# prepare data
./local/preprocess.sh ${conf_path} || exit -1
fi
if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
# train model, all `ckpt` under `train_output_path/checkpoints/` dir
CUDA_VISIBLE_DEVICES=${gpus} ./local/train.sh ${conf_path} ${train_output_path} || exit -1
fi
if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
# synthesize
CUDA_VISIBLE_DEVICES=${gpus} ./local/synthesize.sh ${conf_path} ${train_output_path} ${ckpt_name} || exit -1
fi

@ -0,0 +1,25 @@
stage=-1
stop_stage=100
TARGET_DIR=${MAIN_ROOT}/dataset
. utils/parse_options.sh || exit -1;
src=$1
mkdir -p data/{dev,test}
if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
# download data, generate manifests
# create data/{dev,test} directory to store the manifest files
python3 ${TARGET_DIR}/voxceleb/voxceleb1.py \
--manifest_prefix="data/manifest" \
--target_dir="${src}"
if [ $? -ne 0 ]; then
echo "Prepare Voxceleb failed. Terminated."
exit 1
fi
mv data/manifest.dev data/dev
mv data/voxceleb1.dev.meta data/dev
mv data/manifest.test data/test
mv data/voxceleb1.test.meta data/test
fi

@ -1,4 +1,5 @@
#!/bin/bash
. ./path.sh
set -e

@ -0,0 +1 @@
../../../utils/

@ -1 +1,5 @@
# Changelog
Date: 2022-2-25, Author: Hui Zhang.
- Refactor architecture.
- dtw distance and mcd style dtw

@ -1,170 +0,0 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import List
import numpy as np
from numpy import ndarray as array
from ..backends import depth_convert
from ..utils import ParameterError
__all__ = [
'depth_augment',
'spect_augment',
'random_crop1d',
'random_crop2d',
'adaptive_spect_augment',
]
def randint(high: int) -> int:
"""Generate one random integer in range [0 high)
This is a helper function for random data augmentaiton
"""
return int(np.random.randint(0, high=high))
def rand() -> float:
"""Generate one floating-point number in range [0 1)
This is a helper function for random data augmentaiton
"""
return float(np.random.rand(1))
def depth_augment(y: array,
choices: List=['int8', 'int16'],
probs: List[float]=[0.5, 0.5]) -> array:
""" Audio depth augmentation
Do audio depth augmentation to simulate the distortion brought by quantization.
"""
assert len(probs) == len(
choices
), 'number of choices {} must be equal to size of probs {}'.format(
len(choices), len(probs))
depth = np.random.choice(choices, p=probs)
src_depth = y.dtype
y1 = depth_convert(y, depth)
y2 = depth_convert(y1, src_depth)
return y2
def adaptive_spect_augment(spect: array, tempo_axis: int=0,
level: float=0.1) -> array:
"""Do adpative spectrogram augmentation
The level of the augmentation is gowern by the paramter level,
ranging from 0 to 1, with 0 represents no augmentation
"""
assert spect.ndim == 2., 'only supports 2d tensor or numpy array'
if tempo_axis == 0:
nt, nf = spect.shape
else:
nf, nt = spect.shape
time_mask_width = int(nt * level * 0.5)
freq_mask_width = int(nf * level * 0.5)
num_time_mask = int(10 * level)
num_freq_mask = int(10 * level)
if tempo_axis == 0:
for _ in range(num_time_mask):
start = randint(nt - time_mask_width)
spect[start:start + time_mask_width, :] = 0
for _ in range(num_freq_mask):
start = randint(nf - freq_mask_width)
spect[:, start:start + freq_mask_width] = 0
else:
for _ in range(num_time_mask):
start = randint(nt - time_mask_width)
spect[:, start:start + time_mask_width] = 0
for _ in range(num_freq_mask):
start = randint(nf - freq_mask_width)
spect[start:start + freq_mask_width, :] = 0
return spect
def spect_augment(spect: array,
tempo_axis: int=0,
max_time_mask: int=3,
max_freq_mask: int=3,
max_time_mask_width: int=30,
max_freq_mask_width: int=20) -> array:
"""Do spectrogram augmentation in both time and freq axis
Reference:
"""
assert spect.ndim == 2., 'only supports 2d tensor or numpy array'
if tempo_axis == 0:
nt, nf = spect.shape
else:
nf, nt = spect.shape
num_time_mask = randint(max_time_mask)
num_freq_mask = randint(max_freq_mask)
time_mask_width = randint(max_time_mask_width)
freq_mask_width = randint(max_freq_mask_width)
if tempo_axis == 0:
for _ in range(num_time_mask):
start = randint(nt - time_mask_width)
spect[start:start + time_mask_width, :] = 0
for _ in range(num_freq_mask):
start = randint(nf - freq_mask_width)
spect[:, start:start + freq_mask_width] = 0
else:
for _ in range(num_time_mask):
start = randint(nt - time_mask_width)
spect[:, start:start + time_mask_width] = 0
for _ in range(num_freq_mask):
start = randint(nf - freq_mask_width)
spect[start:start + freq_mask_width, :] = 0
return spect
def random_crop1d(y: array, crop_len: int) -> array:
""" Do random cropping on 1d input signal
The input is a 1d signal, typically a sound waveform
"""
if y.ndim != 1:
'only accept 1d tensor or numpy array'
n = len(y)
idx = randint(n - crop_len)
return y[idx:idx + crop_len]
def random_crop2d(s: array, crop_len: int, tempo_axis: int=0) -> array:
""" Do random cropping for 2D array, typically a spectrogram.
The cropping is done in temporal direction on the time-freq input signal.
"""
if tempo_axis >= s.ndim:
raise ParameterError('axis out of range')
n = s.shape[tempo_axis]
idx = randint(high=n - crop_len)
sli = [slice(None) for i in range(s.ndim)]
sli[tempo_axis] = slice(idx, idx + crop_len)
out = s[tuple(sli)]
return out

@ -1,461 +0,0 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
from functools import partial
from typing import Optional
from typing import Union
import paddle
import paddle.nn as nn
from .window import get_window
__all__ = [
'Spectrogram',
'MelSpectrogram',
'LogMelSpectrogram',
]
def hz_to_mel(freq: Union[paddle.Tensor, float],
htk: bool=False) -> Union[paddle.Tensor, float]:
"""Convert Hz to Mels.
Parameters:
freq: the input tensor of arbitrary shape, or a single floating point number.
htk: use HTK formula to do the conversion.
The default value is False.
Returns:
The frequencies represented in Mel-scale.
"""
if htk:
if isinstance(freq, paddle.Tensor):
return 2595.0 * paddle.log10(1.0 + freq / 700.0)
else:
return 2595.0 * math.log10(1.0 + freq / 700.0)
# Fill in the linear part
f_min = 0.0
f_sp = 200.0 / 3
mels = (freq - f_min) / f_sp
# Fill in the log-scale part
min_log_hz = 1000.0 # beginning of log region (Hz)
min_log_mel = (min_log_hz - f_min) / f_sp # same (Mels)
logstep = math.log(6.4) / 27.0 # step size for log region
if isinstance(freq, paddle.Tensor):
target = min_log_mel + paddle.log(
freq / min_log_hz + 1e-10) / logstep # prevent nan with 1e-10
mask = (freq > min_log_hz).astype(freq.dtype)
mels = target * mask + mels * (
1 - mask) # will replace by masked_fill OP in future
else:
if freq >= min_log_hz:
mels = min_log_mel + math.log(freq / min_log_hz + 1e-10) / logstep
return mels
def mel_to_hz(mel: Union[float, paddle.Tensor],
htk: bool=False) -> Union[float, paddle.Tensor]:
"""Convert mel bin numbers to frequencies.
Parameters:
mel: the mel frequency represented as a tensor of arbitrary shape, or a floating point number.
htk: use HTK formula to do the conversion.
Returns:
The frequencies represented in hz.
"""
if htk:
return 700.0 * (10.0**(mel / 2595.0) - 1.0)
f_min = 0.0
f_sp = 200.0 / 3
freqs = f_min + f_sp * mel
# And now the nonlinear scale
min_log_hz = 1000.0 # beginning of log region (Hz)
min_log_mel = (min_log_hz - f_min) / f_sp # same (Mels)
logstep = math.log(6.4) / 27.0 # step size for log region
if isinstance(mel, paddle.Tensor):
target = min_log_hz * paddle.exp(logstep * (mel - min_log_mel))
mask = (mel > min_log_mel).astype(mel.dtype)
freqs = target * mask + freqs * (
1 - mask) # will replace by masked_fill OP in future
else:
if mel >= min_log_mel:
freqs = min_log_hz * math.exp(logstep * (mel - min_log_mel))
return freqs
def mel_frequencies(n_mels: int=64,
f_min: float=0.0,
f_max: float=11025.0,
htk: bool=False,
dtype: str=paddle.float32):
"""Compute mel frequencies.
Parameters:
n_mels(int): number of Mel bins.
f_min(float): the lower cut-off frequency, below which the filter response is zero.
f_max(float): the upper cut-off frequency, above which the filter response is zero.
htk(bool): whether to use htk formula.
dtype(str): the datatype of the return frequencies.
Returns:
The frequencies represented in Mel-scale
"""
# 'Center freqs' of mel bands - uniformly spaced between limits
min_mel = hz_to_mel(f_min, htk=htk)
max_mel = hz_to_mel(f_max, htk=htk)
mels = paddle.linspace(min_mel, max_mel, n_mels, dtype=dtype)
freqs = mel_to_hz(mels, htk=htk)
return freqs
def fft_frequencies(sr: int, n_fft: int, dtype: str=paddle.float32):
"""Compute fourier frequencies.
Parameters:
sr(int): the audio sample rate.
n_fft(float): the number of fft bins.
dtype(str): the datatype of the return frequencies.
Returns:
The frequencies represented in hz.
"""
return paddle.linspace(0, float(sr) / 2, int(1 + n_fft // 2), dtype=dtype)
def compute_fbank_matrix(sr: int,
n_fft: int,
n_mels: int=64,
f_min: float=0.0,
f_max: Optional[float]=None,
htk: bool=False,
norm: Union[str, float]='slaney',
dtype: str=paddle.float32):
"""Compute fbank matrix.
Parameters:
sr(int): the audio sample rate.
n_fft(int): the number of fft bins.
n_mels(int): the number of Mel bins.
f_min(float): the lower cut-off frequency, below which the filter response is zero.
f_max(float): the upper cut-off frequency, above which the filter response is zero.
htk: whether to use htk formula.
return_complex(bool): whether to return complex matrix. If True, the matrix will
be complex type. Otherwise, the real and image part will be stored in the last
axis of returned tensor.
dtype(str): the datatype of the returned fbank matrix.
Returns:
The fbank matrix of shape (n_mels, int(1+n_fft//2)).
Shape:
output: (n_mels, int(1+n_fft//2))
"""
if f_max is None:
f_max = float(sr) / 2
# Initialize the weights
weights = paddle.zeros((n_mels, int(1 + n_fft // 2)), dtype=dtype)
# Center freqs of each FFT bin
fftfreqs = fft_frequencies(sr=sr, n_fft=n_fft, dtype=dtype)
# 'Center freqs' of mel bands - uniformly spaced between limits
mel_f = mel_frequencies(
n_mels + 2, f_min=f_min, f_max=f_max, htk=htk, dtype=dtype)
fdiff = mel_f[1:] - mel_f[:-1] #np.diff(mel_f)
ramps = mel_f.unsqueeze(1) - fftfreqs.unsqueeze(0)
#ramps = np.subtract.outer(mel_f, fftfreqs)
for i in range(n_mels):
# lower and upper slopes for all bins
lower = -ramps[i] / fdiff[i]
upper = ramps[i + 2] / fdiff[i + 1]
# .. then intersect them with each other and zero
weights[i] = paddle.maximum(
paddle.zeros_like(lower), paddle.minimum(lower, upper))
# Slaney-style mel is scaled to be approx constant energy per channel
if norm == 'slaney':
enorm = 2.0 / (mel_f[2:n_mels + 2] - mel_f[:n_mels])
weights *= enorm.unsqueeze(1)
elif isinstance(norm, int) or isinstance(norm, float):
weights = paddle.nn.functional.normalize(weights, p=norm, axis=-1)
return weights
def power_to_db(magnitude: paddle.Tensor,
ref_value: float=1.0,
amin: float=1e-10,
top_db: Optional[float]=None) -> paddle.Tensor:
"""Convert a power spectrogram (amplitude squared) to decibel (dB) units.
The function computes the scaling ``10 * log10(x / ref)`` in a numerically
stable way.
Parameters:
magnitude(Tensor): the input magnitude tensor of any shape.
ref_value(float): the reference value. If smaller than 1.0, the db level
of the signal will be pulled up accordingly. Otherwise, the db level
is pushed down.
amin(float): the minimum value of input magnitude, below which the input
magnitude is clipped(to amin).
top_db(float): the maximum db value of resulting spectrum, above which the
spectrum is clipped(to top_db).
Returns:
The spectrogram in log-scale.
shape:
input: any shape
output: same as input
"""
if amin <= 0:
raise Exception("amin must be strictly positive")
if ref_value <= 0:
raise Exception("ref_value must be strictly positive")
ones = paddle.ones_like(magnitude)
log_spec = 10.0 * paddle.log10(paddle.maximum(ones * amin, magnitude))
log_spec -= 10.0 * math.log10(max(ref_value, amin))
if top_db is not None:
if top_db < 0:
raise Exception("top_db must be non-negative")
log_spec = paddle.maximum(log_spec, ones * (log_spec.max() - top_db))
return log_spec
class Spectrogram(nn.Layer):
def __init__(self,
n_fft: int=512,
hop_length: Optional[int]=None,
win_length: Optional[int]=None,
window: str='hann',
center: bool=True,
pad_mode: str='reflect',
dtype: str=paddle.float32):
"""Compute spectrogram of a given signal, typically an audio waveform.
The spectorgram is defined as the complex norm of the short-time
Fourier transformation.
Parameters:
n_fft(int): the number of frequency components of the discrete Fourier transform.
The default value is 2048,
hop_length(int|None): the hop length of the short time FFT. If None, it is set to win_length//4.
The default value is None.
win_length: the window length of the short time FFt. If None, it is set to same as n_fft.
The default value is None.
window(str): the name of the window function applied to the single before the Fourier transform.
The folllowing window names are supported: 'hamming','hann','kaiser','gaussian',
'exponential','triang','bohman','blackman','cosine','tukey','taylor'.
The default value is 'hann'
center(bool): if True, the signal is padded so that frame t is centered at x[t * hop_length].
If False, frame t begins at x[t * hop_length]
The default value is True
pad_mode(str): the mode to pad the signal if necessary. The supported modes are 'reflect'
and 'constant'. The default value is 'reflect'.
dtype(str): the data type of input and window.
Notes:
The Spectrogram transform relies on STFT transform to compute the spectrogram.
By default, the weights are not learnable. To fine-tune the Fourier coefficients,
set stop_gradient=False before training.
For more information, see STFT().
"""
super(Spectrogram, self).__init__()
if win_length is None:
win_length = n_fft
fft_window = get_window(window, win_length, fftbins=True, dtype=dtype)
self._stft = partial(
paddle.signal.stft,
n_fft=n_fft,
hop_length=hop_length,
win_length=win_length,
window=fft_window,
center=center,
pad_mode=pad_mode)
def forward(self, x):
stft = self._stft(x)
spectrogram = paddle.square(paddle.abs(stft))
return spectrogram
class MelSpectrogram(nn.Layer):
def __init__(self,
sr: int=22050,
n_fft: int=512,
hop_length: Optional[int]=None,
win_length: Optional[int]=None,
window: str='hann',
center: bool=True,
pad_mode: str='reflect',
n_mels: int=64,
f_min: float=50.0,
f_max: Optional[float]=None,
htk: bool=False,
norm: Union[str, float]='slaney',
dtype: str=paddle.float32):
"""Compute the melspectrogram of a given signal, typically an audio waveform.
The melspectrogram is also known as filterbank or fbank feature in audio community.
It is computed by multiplying spectrogram with Mel filter bank matrix.
Parameters:
sr(int): the audio sample rate.
The default value is 22050.
n_fft(int): the number of frequency components of the discrete Fourier transform.
The default value is 2048,
hop_length(int|None): the hop length of the short time FFT. If None, it is set to win_length//4.
The default value is None.
win_length: the window length of the short time FFt. If None, it is set to same as n_fft.
The default value is None.
window(str): the name of the window function applied to the single before the Fourier transform.
The folllowing window names are supported: 'hamming','hann','kaiser','gaussian',
'exponential','triang','bohman','blackman','cosine','tukey','taylor'.
The default value is 'hann'
center(bool): if True, the signal is padded so that frame t is centered at x[t * hop_length].
If False, frame t begins at x[t * hop_length]
The default value is True
pad_mode(str): the mode to pad the signal if necessary. The supported modes are 'reflect'
and 'constant'.
The default value is 'reflect'.
n_mels(int): the mel bins.
f_min(float): the lower cut-off frequency, below which the filter response is zero.
f_max(float): the upper cut-off frequency, above which the filter response is zeros.
htk(bool): whether to use HTK formula in computing fbank matrix.
norm(str|float): the normalization type in computing fbank matrix. Slaney-style is used by default.
You can specify norm=1.0/2.0 to use customized p-norm normalization.
dtype(str): the datatype of fbank matrix used in the transform. Use float64 to increase numerical
accuracy. Note that the final transform will be conducted in float32 regardless of dtype of fbank matrix.
"""
super(MelSpectrogram, self).__init__()
self._spectrogram = Spectrogram(
n_fft=n_fft,
hop_length=hop_length,
win_length=win_length,
window=window,
center=center,
pad_mode=pad_mode,
dtype=dtype)
self.n_mels = n_mels
self.f_min = f_min
self.f_max = f_max
self.htk = htk
self.norm = norm
if f_max is None:
f_max = sr // 2
self.fbank_matrix = compute_fbank_matrix(
sr=sr,
n_fft=n_fft,
n_mels=n_mels,
f_min=f_min,
f_max=f_max,
htk=htk,
norm=norm,
dtype=dtype) # float64 for better numerical results
self.register_buffer('fbank_matrix', self.fbank_matrix)
def forward(self, x):
spect_feature = self._spectrogram(x)
mel_feature = paddle.matmul(self.fbank_matrix, spect_feature)
return mel_feature
class LogMelSpectrogram(nn.Layer):
def __init__(self,
sr: int=22050,
n_fft: int=512,
hop_length: Optional[int]=None,
win_length: Optional[int]=None,
window: str='hann',
center: bool=True,
pad_mode: str='reflect',
n_mels: int=64,
f_min: float=50.0,
f_max: Optional[float]=None,
htk: bool=False,
norm: Union[str, float]='slaney',
ref_value: float=1.0,
amin: float=1e-10,
top_db: Optional[float]=None,
dtype: str=paddle.float32):
"""Compute log-mel-spectrogram(also known as LogFBank) feature of a given signal,
typically an audio waveform.
Parameters:
sr(int): the audio sample rate.
The default value is 22050.
n_fft(int): the number of frequency components of the discrete Fourier transform.
The default value is 2048,
hop_length(int|None): the hop length of the short time FFT. If None, it is set to win_length//4.
The default value is None.
win_length: the window length of the short time FFt. If None, it is set to same as n_fft.
The default value is None.
window(str): the name of the window function applied to the single before the Fourier transform.
The folllowing window names are supported: 'hamming','hann','kaiser','gaussian',
'exponential','triang','bohman','blackman','cosine','tukey','taylor'.
The default value is 'hann'
center(bool): if True, the signal is padded so that frame t is centered at x[t * hop_length].
If False, frame t begins at x[t * hop_length]
The default value is True
pad_mode(str): the mode to pad the signal if necessary. The supported modes are 'reflect'
and 'constant'.
The default value is 'reflect'.
n_mels(int): the mel bins.
f_min(float): the lower cut-off frequency, below which the filter response is zero.
f_max(float): the upper cut-off frequency, above which the filter response is zeros.
ref_value(float): the reference value. If smaller than 1.0, the db level
htk(bool): whether to use HTK formula in computing fbank matrix.
norm(str|float): the normalization type in computing fbank matrix. Slaney-style is used by default.
You can specify norm=1.0/2.0 to use customized p-norm normalization.
dtype(str): the datatype of fbank matrix used in the transform. Use float64 to increase numerical
accuracy. Note that the final transform will be conducted in float32 regardless of dtype of fbank matrix.
amin(float): the minimum value of input magnitude, below which the input of the signal will be pulled up accordingly.
Otherwise, the db level is pushed down.
magnitude is clipped(to amin). For numerical stability, set amin to a larger value,
e.g., 1e-3.
top_db(float): the maximum db value of resulting spectrum, above which the
spectrum is clipped(to top_db).
"""
super(LogMelSpectrogram, self).__init__()
self._melspectrogram = MelSpectrogram(
sr=sr,
n_fft=n_fft,
hop_length=hop_length,
win_length=win_length,
window=window,
center=center,
pad_mode=pad_mode,
n_mels=n_mels,
f_min=f_min,
f_max=f_max,
htk=htk,
norm=norm,
dtype=dtype)
self.ref_value = ref_value
self.amin = amin
self.top_db = top_db
def forward(self, x):
# import ipdb; ipdb.set_trace()
mel_feature = self._melspectrogram(x)
log_mel_feature = power_to_db(
mel_feature,
ref_value=self.ref_value,
amin=self.amin,
top_db=self.top_db)
return log_mel_feature

@ -0,0 +1,22 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from . import compliance
from . import datasets
from . import features
from . import functional
from . import io
from . import metric
from . import sox_effects
from .backends import load
from .backends import save

@ -0,0 +1,19 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .soundfile_backend import depth_convert
from .soundfile_backend import load
from .soundfile_backend import normalize
from .soundfile_backend import resample
from .soundfile_backend import save
from .soundfile_backend import to_mono

@ -1,4 +1,4 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
@ -29,7 +29,7 @@ __all__ = [
'to_mono',
'depth_convert',
'normalize',
'save_wav',
'save',
'load',
]
NORMALMIZE_TYPES = ['linear', 'gaussian']
@ -41,12 +41,9 @@ EPS = 1e-8
def resample(y: array, src_sr: int, target_sr: int,
mode: str='kaiser_fast') -> array:
""" Audio resampling
This function is the same as using resampy.resample().
Notes:
The default mode is kaiser_fast. For better audio quality, use mode = 'kaiser_fast'
"""
if mode == 'kaiser_best':
@ -106,7 +103,6 @@ def to_mono(y: array, merge_type: str='average') -> array:
def _safe_cast(y: array, dtype: Union[type, str]) -> array:
""" data type casting in a safe way, i.e., prevent overflow or underflow
This function is used internally.
"""
return np.clip(y, np.iinfo(dtype).min, np.iinfo(dtype).max).astype(dtype)
@ -115,10 +111,8 @@ def _safe_cast(y: array, dtype: Union[type, str]) -> array:
def depth_convert(y: array, dtype: Union[type, str],
dithering: bool=True) -> array:
"""Convert audio array to target dtype safely
This function convert audio waveform to a target dtype, with addition steps of
preventing overflow/underflow and preserving audio range.
"""
SUPPORT_DTYPE = ['int16', 'int8', 'float32', 'float64']
@ -168,12 +162,9 @@ def sound_file_load(file: str,
dtype: str='int16',
duration: Optional[int]=None) -> Tuple[array, int]:
"""Load audio using soundfile library
This function load audio file using libsndfile.
Reference:
http://www.mega-nerd.com/libsndfile/#Features
"""
with sf.SoundFile(file) as sf_desc:
sr_native = sf_desc.samplerate
@ -188,33 +179,9 @@ def sound_file_load(file: str,
return y, sf_desc.samplerate
def audio_file_load():
"""Load audio using audiofile library
This function load audio file using audiofile.
Reference:
https://audiofile.68k.org/
"""
raise NotImplementedError()
def sox_file_load():
"""Load audio using sox library
This function load audio file using sox.
Reference:
http://sox.sourceforge.net/
"""
raise NotImplementedError()
def normalize(y: array, norm_type: str='linear',
mul_factor: float=1.0) -> array:
""" normalize an input audio with additional multiplier.
"""
if norm_type == 'linear':
@ -232,14 +199,12 @@ def normalize(y: array, norm_type: str='linear',
return y
def save_wav(y: array, sr: int, file: str) -> None:
def save(y: array, sr: int, file: str) -> None:
"""Save audio file to disk.
This function saves audio to disk using scipy.io.wavfile, with additional step
to convert input waveform to int16 unless it already is int16
Notes:
It only support raw wav format.
"""
if not file.endswith('.wav'):
raise ParameterError(
@ -274,11 +239,8 @@ def load(
resample_mode: str='kaiser_fast') -> Tuple[array, int]:
"""Load audio file from disk.
This function loads audio from disk using using audio beackend.
Parameters:
Notes:
"""
y, r = sound_file_load(file, offset=offset, dtype=dtype, duration=duration)

@ -0,0 +1,13 @@
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

@ -1,4 +1,4 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
@ -11,5 +11,3 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .backends import *
from .features import *

@ -0,0 +1,638 @@
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Modified from torchaudio(https://github.com/pytorch/audio)
import math
from typing import Tuple
import paddle
from paddle import Tensor
from ..functional import create_dct
from ..functional.window import get_window
__all__ = [
'spectrogram',
'fbank',
'mfcc',
]
# window types
HANNING = 'hann'
HAMMING = 'hamming'
POVEY = 'povey'
RECTANGULAR = 'rect'
BLACKMAN = 'blackman'
def _get_epsilon(dtype):
return paddle.to_tensor(1e-07, dtype=dtype)
def _next_power_of_2(x: int) -> int:
return 1 if x == 0 else 2**(x - 1).bit_length()
def _get_strided(waveform: Tensor,
window_size: int,
window_shift: int,
snip_edges: bool) -> Tensor:
assert waveform.dim() == 1
num_samples = waveform.shape[0]
if snip_edges:
if num_samples < window_size:
return paddle.empty((0, 0), dtype=waveform.dtype)
else:
m = 1 + (num_samples - window_size) // window_shift
else:
reversed_waveform = paddle.flip(waveform, [0])
m = (num_samples + (window_shift // 2)) // window_shift
pad = window_size // 2 - window_shift // 2
pad_right = reversed_waveform
if pad > 0:
pad_left = reversed_waveform[-pad:]
waveform = paddle.concat((pad_left, waveform, pad_right), axis=0)
else:
waveform = paddle.concat((waveform[-pad:], pad_right), axis=0)
return paddle.signal.frame(waveform, window_size, window_shift)[:, :m].T
def _feature_window_function(
window_type: str,
window_size: int,
blackman_coeff: float,
dtype: int, ) -> Tensor:
if window_type == HANNING:
return get_window('hann', window_size, fftbins=False, dtype=dtype)
elif window_type == HAMMING:
return get_window('hamming', window_size, fftbins=False, dtype=dtype)
elif window_type == POVEY:
return get_window(
'hann', window_size, fftbins=False, dtype=dtype).pow(0.85)
elif window_type == RECTANGULAR:
return paddle.ones([window_size], dtype=dtype)
elif window_type == BLACKMAN:
a = 2 * math.pi / (window_size - 1)
window_function = paddle.arange(window_size, dtype=dtype)
return (blackman_coeff - 0.5 * paddle.cos(a * window_function) +
(0.5 - blackman_coeff) * paddle.cos(2 * a * window_function)
).astype(dtype)
else:
raise Exception('Invalid window type ' + window_type)
def _get_log_energy(strided_input: Tensor, epsilon: Tensor,
energy_floor: float) -> Tensor:
log_energy = paddle.maximum(strided_input.pow(2).sum(1), epsilon).log()
if energy_floor == 0.0:
return log_energy
return paddle.maximum(
log_energy,
paddle.to_tensor(math.log(energy_floor), dtype=strided_input.dtype))
def _get_waveform_and_window_properties(
waveform: Tensor,
channel: int,
sr: int,
frame_shift: float,
frame_length: float,
round_to_power_of_two: bool,
preemphasis_coefficient: float) -> Tuple[Tensor, int, int, int]:
channel = max(channel, 0)
assert channel < waveform.shape[0], (
'Invalid channel {} for size {}'.format(channel, waveform.shape[0]))
waveform = waveform[channel, :] # size (n)
window_shift = int(
sr * frame_shift *
0.001) # pass frame_shift and frame_length in milliseconds
window_size = int(sr * frame_length * 0.001)
padded_window_size = _next_power_of_2(
window_size) if round_to_power_of_two else window_size
assert 2 <= window_size <= len(waveform), (
'choose a window size {} that is [2, {}]'.format(window_size,
len(waveform)))
assert 0 < window_shift, '`window_shift` must be greater than 0'
assert padded_window_size % 2 == 0, 'the padded `window_size` must be divisible by two.' \
' use `round_to_power_of_two` or change `frame_length`'
assert 0. <= preemphasis_coefficient <= 1.0, '`preemphasis_coefficient` must be between [0,1]'
assert sr > 0, '`sr` must be greater than zero'
return waveform, window_shift, window_size, padded_window_size
def _get_window(waveform: Tensor,
padded_window_size: int,
window_size: int,
window_shift: int,
window_type: str,
blackman_coeff: float,
snip_edges: bool,
raw_energy: bool,
energy_floor: float,
dither: float,
remove_dc_offset: bool,
preemphasis_coefficient: float) -> Tuple[Tensor, Tensor]:
dtype = waveform.dtype
epsilon = _get_epsilon(dtype)
# (m, window_size)
strided_input = _get_strided(waveform, window_size, window_shift,
snip_edges)
if dither != 0.0:
x = paddle.maximum(epsilon,
paddle.rand(strided_input.shape, dtype=dtype))
rand_gauss = paddle.sqrt(-2 * x.log()) * paddle.cos(2 * math.pi * x)
strided_input = strided_input + rand_gauss * dither
if remove_dc_offset:
row_means = paddle.mean(strided_input, axis=1).unsqueeze(1) # (m, 1)
strided_input = strided_input - row_means
if raw_energy:
signal_log_energy = _get_log_energy(strided_input, epsilon,
energy_floor) # (m)
if preemphasis_coefficient != 0.0:
offset_strided_input = paddle.nn.functional.pad(
strided_input.unsqueeze(0), (1, 0),
data_format='NCL',
mode='replicate').squeeze(0) # (m, window_size + 1)
strided_input = strided_input - preemphasis_coefficient * offset_strided_input[:, :
-1]
window_function = _feature_window_function(
window_type, window_size, blackman_coeff,
dtype).unsqueeze(0) # (1, window_size)
strided_input = strided_input * window_function # (m, window_size)
# (m, padded_window_size)
if padded_window_size != window_size:
padding_right = padded_window_size - window_size
strided_input = paddle.nn.functional.pad(
strided_input.unsqueeze(0), (0, padding_right),
data_format='NCL',
mode='constant',
value=0).squeeze(0)
if not raw_energy:
signal_log_energy = _get_log_energy(strided_input, epsilon,
energy_floor) # size (m)
return strided_input, signal_log_energy
def _subtract_column_mean(tensor: Tensor, subtract_mean: bool) -> Tensor:
if subtract_mean:
col_means = paddle.mean(tensor, axis=0).unsqueeze(0)
tensor = tensor - col_means
return tensor
def spectrogram(waveform: Tensor,
blackman_coeff: float=0.42,
channel: int=-1,
dither: float=0.0,
energy_floor: float=1.0,
frame_length: float=25.0,
frame_shift: float=10.0,
preemphasis_coefficient: float=0.97,
raw_energy: bool=True,
remove_dc_offset: bool=True,
round_to_power_of_two: bool=True,
sr: int=16000,
snip_edges: bool=True,
subtract_mean: bool=False,
window_type: str=POVEY) -> Tensor:
"""Compute and return a spectrogram from a waveform. The output is identical to Kaldi's.
Args:
waveform (Tensor): A waveform tensor with shape [C, T].
blackman_coeff (float, optional): Coefficient for Blackman window.. Defaults to 0.42.
channel (int, optional): Select the channel of waveform. Defaults to -1.
dither (float, optional): Dithering constant . Defaults to 0.0.
energy_floor (float, optional): Floor on energy of the output Spectrogram. Defaults to 1.0.
frame_length (float, optional): Frame length in milliseconds. Defaults to 25.0.
frame_shift (float, optional): Shift between adjacent frames in milliseconds. Defaults to 10.0.
preemphasis_coefficient (float, optional): Preemphasis coefficient for input waveform. Defaults to 0.97.
raw_energy (bool, optional): Whether to compute before preemphasis and windowing. Defaults to True.
remove_dc_offset (bool, optional): Whether to subtract mean from waveform on frames. Defaults to True.
round_to_power_of_two (bool, optional): If True, round window size to power of two by zero-padding input
to FFT. Defaults to True.
sr (int, optional): Sample rate of input waveform. Defaults to 16000.
snip_edges (bool, optional): Drop samples in the end of waveform that cann't fit a singal frame when it
is set True. Otherwise performs reflect padding to the end of waveform. Defaults to True.
subtract_mean (bool, optional): Whether to subtract mean of feature files. Defaults to False.
window_type (str, optional): Choose type of window for FFT computation. Defaults to POVEY.
Returns:
Tensor: A spectrogram tensor with shape (m, padded_window_size // 2 + 1) where m is the number of frames
depends on frame_length and frame_shift.
"""
dtype = waveform.dtype
epsilon = _get_epsilon(dtype)
waveform, window_shift, window_size, padded_window_size = _get_waveform_and_window_properties(
waveform, channel, sr, frame_shift, frame_length, round_to_power_of_two,
preemphasis_coefficient)
strided_input, signal_log_energy = _get_window(
waveform, padded_window_size, window_size, window_shift, window_type,
blackman_coeff, snip_edges, raw_energy, energy_floor, dither,
remove_dc_offset, preemphasis_coefficient)
# (m, padded_window_size // 2 + 1, 2)
fft = paddle.fft.rfft(strided_input)
power_spectrum = paddle.maximum(
fft.abs().pow(2.), epsilon).log() # (m, padded_window_size // 2 + 1)
power_spectrum[:, 0] = signal_log_energy
power_spectrum = _subtract_column_mean(power_spectrum, subtract_mean)
return power_spectrum
def _inverse_mel_scale_scalar(mel_freq: float) -> float:
return 700.0 * (math.exp(mel_freq / 1127.0) - 1.0)
def _inverse_mel_scale(mel_freq: Tensor) -> Tensor:
return 700.0 * ((mel_freq / 1127.0).exp() - 1.0)
def _mel_scale_scalar(freq: float) -> float:
return 1127.0 * math.log(1.0 + freq / 700.0)
def _mel_scale(freq: Tensor) -> Tensor:
return 1127.0 * (1.0 + freq / 700.0).log()
def _vtln_warp_freq(vtln_low_cutoff: float,
vtln_high_cutoff: float,
low_freq: float,
high_freq: float,
vtln_warp_factor: float,
freq: Tensor) -> Tensor:
assert vtln_low_cutoff > low_freq, 'be sure to set the vtln_low option higher than low_freq'
assert vtln_high_cutoff < high_freq, 'be sure to set the vtln_high option lower than high_freq [or negative]'
l = vtln_low_cutoff * max(1.0, vtln_warp_factor)
h = vtln_high_cutoff * min(1.0, vtln_warp_factor)
scale = 1.0 / vtln_warp_factor
Fl = scale * l
Fh = scale * h
assert l > low_freq and h < high_freq
scale_left = (Fl - low_freq) / (l - low_freq)
scale_right = (high_freq - Fh) / (high_freq - h)
res = paddle.empty_like(freq)
outside_low_high_freq = paddle.less_than(freq, paddle.to_tensor(low_freq)) \
| paddle.greater_than(freq, paddle.to_tensor(high_freq))
before_l = paddle.less_than(freq, paddle.to_tensor(l))
before_h = paddle.less_than(freq, paddle.to_tensor(h))
after_h = paddle.greater_equal(freq, paddle.to_tensor(h))
res[after_h] = high_freq + scale_right * (freq[after_h] - high_freq)
res[before_h] = scale * freq[before_h]
res[before_l] = low_freq + scale_left * (freq[before_l] - low_freq)
res[outside_low_high_freq] = freq[outside_low_high_freq]
return res
def _vtln_warp_mel_freq(vtln_low_cutoff: float,
vtln_high_cutoff: float,
low_freq,
high_freq: float,
vtln_warp_factor: float,
mel_freq: Tensor) -> Tensor:
return _mel_scale(
_vtln_warp_freq(vtln_low_cutoff, vtln_high_cutoff, low_freq, high_freq,
vtln_warp_factor, _inverse_mel_scale(mel_freq)))
def _get_mel_banks(num_bins: int,
window_length_padded: int,
sample_freq: float,
low_freq: float,
high_freq: float,
vtln_low: float,
vtln_high: float,
vtln_warp_factor: float) -> Tuple[Tensor, Tensor]:
assert num_bins > 3, 'Must have at least 3 mel bins'
assert window_length_padded % 2 == 0
num_fft_bins = window_length_padded / 2
nyquist = 0.5 * sample_freq
if high_freq <= 0.0:
high_freq += nyquist
assert (0.0 <= low_freq < nyquist) and (0.0 < high_freq <= nyquist) and (low_freq < high_freq), \
('Bad values in options: low-freq {} and high-freq {} vs. nyquist {}'.format(low_freq, high_freq, nyquist))
fft_bin_width = sample_freq / window_length_padded
mel_low_freq = _mel_scale_scalar(low_freq)
mel_high_freq = _mel_scale_scalar(high_freq)
mel_freq_delta = (mel_high_freq - mel_low_freq) / (num_bins + 1)
if vtln_high < 0.0:
vtln_high += nyquist
assert vtln_warp_factor == 1.0 or ((low_freq < vtln_low < high_freq) and
(0.0 < vtln_high < high_freq) and (vtln_low < vtln_high)), \
('Bad values in options: vtln-low {} and vtln-high {}, versus '
'low-freq {} and high-freq {}'.format(vtln_low, vtln_high, low_freq, high_freq))
bin = paddle.arange(num_bins).unsqueeze(1)
left_mel = mel_low_freq + bin * mel_freq_delta # (num_bins, 1)
center_mel = mel_low_freq + (bin + 1.0) * mel_freq_delta # (num_bins, 1)
right_mel = mel_low_freq + (bin + 2.0) * mel_freq_delta # (num_bins, 1)
if vtln_warp_factor != 1.0:
left_mel = _vtln_warp_mel_freq(vtln_low, vtln_high, low_freq, high_freq,
vtln_warp_factor, left_mel)
center_mel = _vtln_warp_mel_freq(vtln_low, vtln_high, low_freq,
high_freq, vtln_warp_factor,
center_mel)
right_mel = _vtln_warp_mel_freq(vtln_low, vtln_high, low_freq,
high_freq, vtln_warp_factor, right_mel)
center_freqs = _inverse_mel_scale(center_mel) # (num_bins)
# (1, num_fft_bins)
mel = _mel_scale(fft_bin_width * paddle.arange(num_fft_bins)).unsqueeze(0)
# (num_bins, num_fft_bins)
up_slope = (mel - left_mel) / (center_mel - left_mel)
down_slope = (right_mel - mel) / (right_mel - center_mel)
if vtln_warp_factor == 1.0:
bins = paddle.maximum(
paddle.zeros([1]), paddle.minimum(up_slope, down_slope))
else:
bins = paddle.zeros_like(up_slope)
up_idx = paddle.greater_than(mel, left_mel) & paddle.less_than(
mel, center_mel)
down_idx = paddle.greater_than(mel, center_mel) & paddle.less_than(
mel, right_mel)
bins[up_idx] = up_slope[up_idx]
bins[down_idx] = down_slope[down_idx]
return bins, center_freqs
def fbank(waveform: Tensor,
blackman_coeff: float=0.42,
channel: int=-1,
dither: float=0.0,
energy_floor: float=1.0,
frame_length: float=25.0,
frame_shift: float=10.0,
high_freq: float=0.0,
htk_compat: bool=False,
low_freq: float=20.0,
n_mels: int=23,
preemphasis_coefficient: float=0.97,
raw_energy: bool=True,
remove_dc_offset: bool=True,
round_to_power_of_two: bool=True,
sr: int=16000,
snip_edges: bool=True,
subtract_mean: bool=False,
use_energy: bool=False,
use_log_fbank: bool=True,
use_power: bool=True,
vtln_high: float=-500.0,
vtln_low: float=100.0,
vtln_warp: float=1.0,
window_type: str=POVEY) -> Tensor:
"""Compute and return filter banks from a waveform. The output is identical to Kaldi's.
Args:
waveform (Tensor): A waveform tensor with shape [C, T].
blackman_coeff (float, optional): Coefficient for Blackman window.. Defaults to 0.42.
channel (int, optional): Select the channel of waveform. Defaults to -1.
dither (float, optional): Dithering constant . Defaults to 0.0.
energy_floor (float, optional): Floor on energy of the output Spectrogram. Defaults to 1.0.
frame_length (float, optional): Frame length in milliseconds. Defaults to 25.0.
frame_shift (float, optional): Shift between adjacent frames in milliseconds. Defaults to 10.0.
high_freq (float, optional): The upper cut-off frequency. Defaults to 0.0.
htk_compat (bool, optional): Put energy to the last when it is set True. Defaults to False.
low_freq (float, optional): The lower cut-off frequency. Defaults to 20.0.
n_mels (int, optional): Number of output mel bins. Defaults to 23.
preemphasis_coefficient (float, optional): Preemphasis coefficient for input waveform. Defaults to 0.97.
raw_energy (bool, optional): Whether to compute before preemphasis and windowing. Defaults to True.
remove_dc_offset (bool, optional): Whether to subtract mean from waveform on frames. Defaults to True.
round_to_power_of_two (bool, optional): If True, round window size to power of two by zero-padding input
to FFT. Defaults to True.
sr (int, optional): Sample rate of input waveform. Defaults to 16000.
snip_edges (bool, optional): Drop samples in the end of waveform that cann't fit a singal frame when it
is set True. Otherwise performs reflect padding to the end of waveform. Defaults to True.
subtract_mean (bool, optional): Whether to subtract mean of feature files. Defaults to False.
use_energy (bool, optional): Add an dimension with energy of spectrogram to the output. Defaults to False.
use_log_fbank (bool, optional): Return log fbank when it is set True. Defaults to True.
use_power (bool, optional): Whether to use power instead of magnitude. Defaults to True.
vtln_high (float, optional): High inflection point in piecewise linear VTLN warping function. Defaults to -500.0.
vtln_low (float, optional): Low inflection point in piecewise linear VTLN warping function. Defaults to 100.0.
vtln_warp (float, optional): Vtln warp factor. Defaults to 1.0.
window_type (str, optional): Choose type of window for FFT computation. Defaults to POVEY.
Returns:
Tensor: A filter banks tensor with shape (m, n_mels).
"""
dtype = waveform.dtype
waveform, window_shift, window_size, padded_window_size = _get_waveform_and_window_properties(
waveform, channel, sr, frame_shift, frame_length, round_to_power_of_two,
preemphasis_coefficient)
strided_input, signal_log_energy = _get_window(
waveform, padded_window_size, window_size, window_shift, window_type,
blackman_coeff, snip_edges, raw_energy, energy_floor, dither,
remove_dc_offset, preemphasis_coefficient)
# (m, padded_window_size // 2 + 1)
spectrum = paddle.fft.rfft(strided_input).abs()
if use_power:
spectrum = spectrum.pow(2.)
# (n_mels, padded_window_size // 2)
mel_energies, _ = _get_mel_banks(n_mels, padded_window_size, sr, low_freq,
high_freq, vtln_low, vtln_high, vtln_warp)
mel_energies = mel_energies.astype(dtype)
# (n_mels, padded_window_size // 2 + 1)
mel_energies = paddle.nn.functional.pad(
mel_energies.unsqueeze(0), (0, 1),
data_format='NCL',
mode='constant',
value=0).squeeze(0)
# (m, n_mels)
mel_energies = paddle.mm(spectrum, mel_energies.T)
if use_log_fbank:
mel_energies = paddle.maximum(mel_energies, _get_epsilon(dtype)).log()
if use_energy:
signal_log_energy = signal_log_energy.unsqueeze(1)
if htk_compat:
mel_energies = paddle.concat(
(mel_energies, signal_log_energy), axis=1)
else:
mel_energies = paddle.concat(
(signal_log_energy, mel_energies), axis=1)
# (m, n_mels + 1)
mel_energies = _subtract_column_mean(mel_energies, subtract_mean)
return mel_energies
def _get_dct_matrix(n_mfcc: int, n_mels: int) -> Tensor:
dct_matrix = create_dct(n_mels, n_mels, 'ortho')
dct_matrix[:, 0] = math.sqrt(1 / float(n_mels))
dct_matrix = dct_matrix[:, :n_mfcc] # (n_mels, n_mfcc)
return dct_matrix
def _get_lifter_coeffs(n_mfcc: int, cepstral_lifter: float) -> Tensor:
i = paddle.arange(n_mfcc)
return 1.0 + 0.5 * cepstral_lifter * paddle.sin(math.pi * i /
cepstral_lifter)
def mfcc(waveform: Tensor,
blackman_coeff: float=0.42,
cepstral_lifter: float=22.0,
channel: int=-1,
dither: float=0.0,
energy_floor: float=1.0,
frame_length: float=25.0,
frame_shift: float=10.0,
high_freq: float=0.0,
htk_compat: bool=False,
low_freq: float=20.0,
n_mfcc: int=13,
n_mels: int=23,
preemphasis_coefficient: float=0.97,
raw_energy: bool=True,
remove_dc_offset: bool=True,
round_to_power_of_two: bool=True,
sr: int=16000,
snip_edges: bool=True,
subtract_mean: bool=False,
use_energy: bool=False,
vtln_high: float=-500.0,
vtln_low: float=100.0,
vtln_warp: float=1.0,
window_type: str=POVEY) -> Tensor:
"""Compute and return mel frequency cepstral coefficients from a waveform. The output is
identical to Kaldi's.
Args:
waveform (Tensor): A waveform tensor with shape [C, T].
blackman_coeff (float, optional): Coefficient for Blackman window.. Defaults to 0.42.
cepstral_lifter (float, optional): Scaling of output mfccs. Defaults to 22.0.
channel (int, optional): Select the channel of waveform. Defaults to -1.
dither (float, optional): Dithering constant . Defaults to 0.0.
energy_floor (float, optional): Floor on energy of the output Spectrogram. Defaults to 1.0.
frame_length (float, optional): Frame length in milliseconds. Defaults to 25.0.
frame_shift (float, optional): Shift between adjacent frames in milliseconds. Defaults to 10.0.
high_freq (float, optional): The upper cut-off frequency. Defaults to 0.0.
htk_compat (bool, optional): Put energy to the last when it is set True. Defaults to False.
low_freq (float, optional): The lower cut-off frequency. Defaults to 20.0.
n_mfcc (int, optional): Number of cepstra in MFCC. Defaults to 13.
n_mels (int, optional): Number of output mel bins. Defaults to 23.
preemphasis_coefficient (float, optional): Preemphasis coefficient for input waveform. Defaults to 0.97.
raw_energy (bool, optional): Whether to compute before preemphasis and windowing. Defaults to True.
remove_dc_offset (bool, optional): Whether to subtract mean from waveform on frames. Defaults to True.
round_to_power_of_two (bool, optional): If True, round window size to power of two by zero-padding input
to FFT. Defaults to True.
sr (int, optional): Sample rate of input waveform. Defaults to 16000.
snip_edges (bool, optional): Drop samples in the end of waveform that cann't fit a singal frame when it
is set True. Otherwise performs reflect padding to the end of waveform. Defaults to True.
subtract_mean (bool, optional): Whether to subtract mean of feature files. Defaults to False.
use_energy (bool, optional): Add an dimension with energy of spectrogram to the output. Defaults to False.
vtln_high (float, optional): High inflection point in piecewise linear VTLN warping function. Defaults to -500.0.
vtln_low (float, optional): Low inflection point in piecewise linear VTLN warping function. Defaults to 100.0.
vtln_warp (float, optional): Vtln warp factor. Defaults to 1.0.
window_type (str, optional): Choose type of window for FFT computation. Defaults to POVEY.
Returns:
Tensor: A mel frequency cepstral coefficients tensor with shape (m, n_mfcc).
"""
assert n_mfcc <= n_mels, 'n_mfcc cannot be larger than n_mels: %d vs %d' % (
n_mfcc, n_mels)
dtype = waveform.dtype
# (m, n_mels + use_energy)
feature = fbank(
waveform=waveform,
blackman_coeff=blackman_coeff,
channel=channel,
dither=dither,
energy_floor=energy_floor,
frame_length=frame_length,
frame_shift=frame_shift,
high_freq=high_freq,
htk_compat=htk_compat,
low_freq=low_freq,
n_mels=n_mels,
preemphasis_coefficient=preemphasis_coefficient,
raw_energy=raw_energy,
remove_dc_offset=remove_dc_offset,
round_to_power_of_two=round_to_power_of_two,
sr=sr,
snip_edges=snip_edges,
subtract_mean=False,
use_energy=use_energy,
use_log_fbank=True,
use_power=True,
vtln_high=vtln_high,
vtln_low=vtln_low,
vtln_warp=vtln_warp,
window_type=window_type)
if use_energy:
# (m)
signal_log_energy = feature[:, n_mels if htk_compat else 0]
mel_offset = int(not htk_compat)
feature = feature[:, mel_offset:(n_mels + mel_offset)]
# (n_mels, n_mfcc)
dct_matrix = _get_dct_matrix(n_mfcc, n_mels).astype(dtype=dtype)
# (m, n_mfcc)
feature = feature.matmul(dct_matrix)
if cepstral_lifter != 0.0:
# (1, n_mfcc)
lifter_coeffs = _get_lifter_coeffs(n_mfcc, cepstral_lifter).unsqueeze(0)
feature *= lifter_coeffs.astype(dtype=dtype)
if use_energy:
feature[:, 0] = signal_log_energy
if htk_compat:
energy = feature[:, 0].unsqueeze(1) # (m, 1)
feature = feature[:, 1:] # (m, n_mfcc - 1)
if not use_energy:
energy *= math.sqrt(2)
feature = paddle.concat((feature, energy), axis=1)
feature = _subtract_column_mean(feature, subtract_mean)
return feature

@ -21,11 +21,13 @@ import numpy as np
import scipy
from numpy import ndarray as array
from numpy.lib.stride_tricks import as_strided
from scipy.signal import get_window
from scipy import signal
from ..backends import depth_convert
from ..utils import ParameterError
__all__ = [
# dsp
'stft',
'mfcc',
'hz_to_mel',
@ -38,6 +40,12 @@ __all__ = [
'spectrogram',
'mu_encode',
'mu_decode',
# augmentation
'depth_augment',
'spect_augment',
'random_crop1d',
'random_crop2d',
'adaptive_spect_augment',
]
@ -303,7 +311,7 @@ def stft(x: array,
if hop_length is None:
hop_length = int(win_length // 4)
fft_window = get_window(window, win_length, fftbins=True)
fft_window = signal.get_window(window, win_length, fftbins=True)
# Pad the window out to n_fft size
fft_window = pad_center(fft_window, n_fft)
@ -576,3 +584,145 @@ def mu_decode(y: array, mu: int=255, quantized: bool=True) -> array:
y = y * 2 / mu - 1
x = np.sign(y) / mu * ((1 + mu)**np.abs(y) - 1)
return x
def randint(high: int) -> int:
"""Generate one random integer in range [0 high)
This is a helper function for random data augmentaiton
"""
return int(np.random.randint(0, high=high))
def rand() -> float:
"""Generate one floating-point number in range [0 1)
This is a helper function for random data augmentaiton
"""
return float(np.random.rand(1))
def depth_augment(y: array,
choices: List=['int8', 'int16'],
probs: List[float]=[0.5, 0.5]) -> array:
""" Audio depth augmentation
Do audio depth augmentation to simulate the distortion brought by quantization.
"""
assert len(probs) == len(
choices
), 'number of choices {} must be equal to size of probs {}'.format(
len(choices), len(probs))
depth = np.random.choice(choices, p=probs)
src_depth = y.dtype
y1 = depth_convert(y, depth)
y2 = depth_convert(y1, src_depth)
return y2
def adaptive_spect_augment(spect: array, tempo_axis: int=0,
level: float=0.1) -> array:
"""Do adpative spectrogram augmentation
The level of the augmentation is gowern by the paramter level,
ranging from 0 to 1, with 0 represents no augmentation
"""
assert spect.ndim == 2., 'only supports 2d tensor or numpy array'
if tempo_axis == 0:
nt, nf = spect.shape
else:
nf, nt = spect.shape
time_mask_width = int(nt * level * 0.5)
freq_mask_width = int(nf * level * 0.5)
num_time_mask = int(10 * level)
num_freq_mask = int(10 * level)
if tempo_axis == 0:
for _ in range(num_time_mask):
start = randint(nt - time_mask_width)
spect[start:start + time_mask_width, :] = 0
for _ in range(num_freq_mask):
start = randint(nf - freq_mask_width)
spect[:, start:start + freq_mask_width] = 0
else:
for _ in range(num_time_mask):
start = randint(nt - time_mask_width)
spect[:, start:start + time_mask_width] = 0
for _ in range(num_freq_mask):
start = randint(nf - freq_mask_width)
spect[start:start + freq_mask_width, :] = 0
return spect
def spect_augment(spect: array,
tempo_axis: int=0,
max_time_mask: int=3,
max_freq_mask: int=3,
max_time_mask_width: int=30,
max_freq_mask_width: int=20) -> array:
"""Do spectrogram augmentation in both time and freq axis
Reference:
"""
assert spect.ndim == 2., 'only supports 2d tensor or numpy array'
if tempo_axis == 0:
nt, nf = spect.shape
else:
nf, nt = spect.shape
num_time_mask = randint(max_time_mask)
num_freq_mask = randint(max_freq_mask)
time_mask_width = randint(max_time_mask_width)
freq_mask_width = randint(max_freq_mask_width)
if tempo_axis == 0:
for _ in range(num_time_mask):
start = randint(nt - time_mask_width)
spect[start:start + time_mask_width, :] = 0
for _ in range(num_freq_mask):
start = randint(nf - freq_mask_width)
spect[:, start:start + freq_mask_width] = 0
else:
for _ in range(num_time_mask):
start = randint(nt - time_mask_width)
spect[:, start:start + time_mask_width] = 0
for _ in range(num_freq_mask):
start = randint(nf - freq_mask_width)
spect[start:start + freq_mask_width, :] = 0
return spect
def random_crop1d(y: array, crop_len: int) -> array:
""" Do random cropping on 1d input signal
The input is a 1d signal, typically a sound waveform
"""
if y.ndim != 1:
'only accept 1d tensor or numpy array'
n = len(y)
idx = randint(n - crop_len)
return y[idx:idx + crop_len]
def random_crop2d(s: array, crop_len: int, tempo_axis: int=0) -> array:
""" Do random cropping for 2D array, typically a spectrogram.
The cropping is done in temporal direction on the time-freq input signal.
"""
if tempo_axis >= s.ndim:
raise ParameterError('axis out of range')
n = s.shape[tempo_axis]
idx = randint(high=n - crop_len)
sli = [slice(None) for i in range(s.ndim)]
sli[tempo_axis] = slice(idx, idx + crop_len)
out = s[tuple(sli)]
return out

@ -15,10 +15,3 @@ from .esc50 import ESC50
from .gtzan import GTZAN
from .tess import TESS
from .urban_sound import UrbanSound8K
__all__ = [
'ESC50',
'UrbanSound8K',
'GTZAN',
'TESS',
]

@ -17,8 +17,8 @@ import numpy as np
import paddle
from ..backends import load as load_audio
from ..features import melspectrogram
from ..features import mfcc
from ..compliance.librosa import melspectrogram
from ..compliance.librosa import mfcc
feat_funcs = {
'raw': None,

@ -11,6 +11,7 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .augment import *
from .core import *
from .spectrum import *
from .layers import LogMelSpectrogram
from .layers import MelSpectrogram
from .layers import MFCC
from .layers import Spectrogram

@ -0,0 +1,344 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from functools import partial
from typing import Optional
from typing import Union
import paddle
import paddle.nn as nn
from ..functional import compute_fbank_matrix
from ..functional import create_dct
from ..functional import power_to_db
from ..functional.window import get_window
__all__ = [
'Spectrogram',
'MelSpectrogram',
'LogMelSpectrogram',
'MFCC',
]
class Spectrogram(nn.Layer):
def __init__(self,
n_fft: int=512,
hop_length: Optional[int]=None,
win_length: Optional[int]=None,
window: str='hann',
center: bool=True,
pad_mode: str='reflect',
dtype: str=paddle.float32):
"""Compute spectrogram of a given signal, typically an audio waveform.
The spectorgram is defined as the complex norm of the short-time
Fourier transformation.
Parameters:
n_fft (int): the number of frequency components of the discrete Fourier transform.
The default value is 2048,
hop_length (int|None): the hop length of the short time FFT. If None, it is set to win_length//4.
The default value is None.
win_length: the window length of the short time FFt. If None, it is set to same as n_fft.
The default value is None.
window (str): the name of the window function applied to the single before the Fourier transform.
The folllowing window names are supported: 'hamming','hann','kaiser','gaussian',
'exponential','triang','bohman','blackman','cosine','tukey','taylor'.
The default value is 'hann'
center (bool): if True, the signal is padded so that frame t is centered at x[t * hop_length].
If False, frame t begins at x[t * hop_length]
The default value is True
pad_mode (str): the mode to pad the signal if necessary. The supported modes are 'reflect'
and 'constant'. The default value is 'reflect'.
dtype (str): the data type of input and window.
Notes:
The Spectrogram transform relies on STFT transform to compute the spectrogram.
By default, the weights are not learnable. To fine-tune the Fourier coefficients,
set stop_gradient=False before training.
For more information, see STFT().
"""
super(Spectrogram, self).__init__()
if win_length is None:
win_length = n_fft
self.fft_window = get_window(
window, win_length, fftbins=True, dtype=dtype)
self._stft = partial(
paddle.signal.stft,
n_fft=n_fft,
hop_length=hop_length,
win_length=win_length,
window=self.fft_window,
center=center,
pad_mode=pad_mode)
self.register_buffer('fft_window', self.fft_window)
def forward(self, x):
stft = self._stft(x)
spectrogram = paddle.square(paddle.abs(stft))
return spectrogram
class MelSpectrogram(nn.Layer):
def __init__(self,
sr: int=22050,
n_fft: int=512,
hop_length: Optional[int]=None,
win_length: Optional[int]=None,
window: str='hann',
center: bool=True,
pad_mode: str='reflect',
n_mels: int=64,
f_min: float=50.0,
f_max: Optional[float]=None,
htk: bool=False,
norm: Union[str, float]='slaney',
dtype: str=paddle.float32):
"""Compute the melspectrogram of a given signal, typically an audio waveform.
The melspectrogram is also known as filterbank or fbank feature in audio community.
It is computed by multiplying spectrogram with Mel filter bank matrix.
Parameters:
sr(int): the audio sample rate.
The default value is 22050.
n_fft(int): the number of frequency components of the discrete Fourier transform.
The default value is 2048,
hop_length(int|None): the hop length of the short time FFT. If None, it is set to win_length//4.
The default value is None.
win_length: the window length of the short time FFt. If None, it is set to same as n_fft.
The default value is None.
window(str): the name of the window function applied to the single before the Fourier transform.
The folllowing window names are supported: 'hamming','hann','kaiser','gaussian',
'exponential','triang','bohman','blackman','cosine','tukey','taylor'.
The default value is 'hann'
center(bool): if True, the signal is padded so that frame t is centered at x[t * hop_length].
If False, frame t begins at x[t * hop_length]
The default value is True
pad_mode(str): the mode to pad the signal if necessary. The supported modes are 'reflect'
and 'constant'.
The default value is 'reflect'.
n_mels(int): the mel bins.
f_min(float): the lower cut-off frequency, below which the filter response is zero.
f_max(float): the upper cut-off frequency, above which the filter response is zeros.
htk(bool): whether to use HTK formula in computing fbank matrix.
norm(str|float): the normalization type in computing fbank matrix. Slaney-style is used by default.
You can specify norm=1.0/2.0 to use customized p-norm normalization.
dtype(str): the datatype of fbank matrix used in the transform. Use float64 to increase numerical
accuracy. Note that the final transform will be conducted in float32 regardless of dtype of fbank matrix.
"""
super(MelSpectrogram, self).__init__()
self._spectrogram = Spectrogram(
n_fft=n_fft,
hop_length=hop_length,
win_length=win_length,
window=window,
center=center,
pad_mode=pad_mode,
dtype=dtype)
self.n_mels = n_mels
self.f_min = f_min
self.f_max = f_max
self.htk = htk
self.norm = norm
if f_max is None:
f_max = sr // 2
self.fbank_matrix = compute_fbank_matrix(
sr=sr,
n_fft=n_fft,
n_mels=n_mels,
f_min=f_min,
f_max=f_max,
htk=htk,
norm=norm,
dtype=dtype) # float64 for better numerical results
self.register_buffer('fbank_matrix', self.fbank_matrix)
def forward(self, x):
spect_feature = self._spectrogram(x)
mel_feature = paddle.matmul(self.fbank_matrix, spect_feature)
return mel_feature
class LogMelSpectrogram(nn.Layer):
def __init__(self,
sr: int=22050,
n_fft: int=512,
hop_length: Optional[int]=None,
win_length: Optional[int]=None,
window: str='hann',
center: bool=True,
pad_mode: str='reflect',
n_mels: int=64,
f_min: float=50.0,
f_max: Optional[float]=None,
htk: bool=False,
norm: Union[str, float]='slaney',
ref_value: float=1.0,
amin: float=1e-10,
top_db: Optional[float]=None,
dtype: str=paddle.float32):
"""Compute log-mel-spectrogram(also known as LogFBank) feature of a given signal,
typically an audio waveform.
Parameters:
sr (int): the audio sample rate.
The default value is 22050.
n_fft (int): the number of frequency components of the discrete Fourier transform.
The default value is 2048,
hop_length (int|None): the hop length of the short time FFT. If None, it is set to win_length//4.
The default value is None.
win_length: the window length of the short time FFt. If None, it is set to same as n_fft.
The default value is None.
window (str): the name of the window function applied to the single before the Fourier transform.
The folllowing window names are supported: 'hamming','hann','kaiser','gaussian',
'exponential','triang','bohman','blackman','cosine','tukey','taylor'.
The default value is 'hann'
center (bool): if True, the signal is padded so that frame t is centered at x[t * hop_length].
If False, frame t begins at x[t * hop_length]
The default value is True
pad_mode (str): the mode to pad the signal if necessary. The supported modes are 'reflect'
and 'constant'.
The default value is 'reflect'.
n_mels (int): the mel bins.
f_min (float): the lower cut-off frequency, below which the filter response is zero.
f_max (float): the upper cut-off frequency, above which the filter response is zeros.
htk (bool): whether to use HTK formula in computing fbank matrix.
norm (str|float): the normalization type in computing fbank matrix. Slaney-style is used by default.
You can specify norm=1.0/2.0 to use customized p-norm normalization.
ref_value (float): the reference value. If smaller than 1.0, the db level
amin (float): the minimum value of input magnitude, below which the input of the signal will be pulled up accordingly.
Otherwise, the db level is pushed down.
magnitude is clipped(to amin). For numerical stability, set amin to a larger value,
e.g., 1e-3.
top_db (float): the maximum db value of resulting spectrum, above which the
spectrum is clipped(to top_db).
dtype (str): the datatype of fbank matrix used in the transform. Use float64 to increase numerical
accuracy. Note that the final transform will be conducted in float32 regardless of dtype of fbank matrix.
"""
super(LogMelSpectrogram, self).__init__()
self._melspectrogram = MelSpectrogram(
sr=sr,
n_fft=n_fft,
hop_length=hop_length,
win_length=win_length,
window=window,
center=center,
pad_mode=pad_mode,
n_mels=n_mels,
f_min=f_min,
f_max=f_max,
htk=htk,
norm=norm,
dtype=dtype)
self.ref_value = ref_value
self.amin = amin
self.top_db = top_db
def forward(self, x):
# import ipdb; ipdb.set_trace()
mel_feature = self._melspectrogram(x)
log_mel_feature = power_to_db(
mel_feature,
ref_value=self.ref_value,
amin=self.amin,
top_db=self.top_db)
return log_mel_feature
class MFCC(nn.Layer):
def __init__(self,
sr: int=22050,
n_mfcc: int=40,
n_fft: int=512,
hop_length: Optional[int]=None,
win_length: Optional[int]=None,
window: str='hann',
center: bool=True,
pad_mode: str='reflect',
n_mels: int=64,
f_min: float=50.0,
f_max: Optional[float]=None,
htk: bool=False,
norm: Union[str, float]='slaney',
ref_value: float=1.0,
amin: float=1e-10,
top_db: Optional[float]=None,
dtype: str=paddle.float32):
"""Compute mel frequency cepstral coefficients(MFCCs) feature of given waveforms.
Parameters:
sr(int): the audio sample rate.
The default value is 22050.
n_mfcc (int, optional): Number of cepstra in MFCC. Defaults to 40.
n_fft (int): the number of frequency components of the discrete Fourier transform.
The default value is 2048,
hop_length (int|None): the hop length of the short time FFT. If None, it is set to win_length//4.
The default value is None.
win_length: the window length of the short time FFt. If None, it is set to same as n_fft.
The default value is None.
window (str): the name of the window function applied to the single before the Fourier transform.
The folllowing window names are supported: 'hamming','hann','kaiser','gaussian',
'exponential','triang','bohman','blackman','cosine','tukey','taylor'.
The default value is 'hann'
center (bool): if True, the signal is padded so that frame t is centered at x[t * hop_length].
If False, frame t begins at x[t * hop_length]
The default value is True
pad_mode (str): the mode to pad the signal if necessary. The supported modes are 'reflect'
and 'constant'.
The default value is 'reflect'.
n_mels (int): the mel bins.
f_min (float): the lower cut-off frequency, below which the filter response is zero.
f_max (float): the upper cut-off frequency, above which the filter response is zeros.
htk (bool): whether to use HTK formula in computing fbank matrix.
norm (str|float): the normalization type in computing fbank matrix. Slaney-style is used by default.
You can specify norm=1.0/2.0 to use customized p-norm normalization.
ref_value (float): the reference value. If smaller than 1.0, the db level
amin (float): the minimum value of input magnitude, below which the input of the signal will be pulled up accordingly.
Otherwise, the db level is pushed down.
magnitude is clipped(to amin). For numerical stability, set amin to a larger value,
e.g., 1e-3.
top_db (float): the maximum db value of resulting spectrum, above which the
spectrum is clipped(to top_db).
dtype (str): the datatype of fbank matrix used in the transform. Use float64 to increase numerical
accuracy. Note that the final transform will be conducted in float32 regardless of dtype of fbank matrix.
"""
super(MFCC, self).__init__()
assert n_mfcc <= n_mels, 'n_mfcc cannot be larger than n_mels: %d vs %d' % (
n_mfcc, n_mels)
self._log_melspectrogram = LogMelSpectrogram(
sr=sr,
n_fft=n_fft,
hop_length=hop_length,
win_length=win_length,
window=window,
center=center,
pad_mode=pad_mode,
n_mels=n_mels,
f_min=f_min,
f_max=f_max,
htk=htk,
norm=norm,
ref_value=ref_value,
amin=amin,
top_db=top_db,
dtype=dtype)
self.dct_matrix = create_dct(n_mfcc=n_mfcc, n_mels=n_mels, dtype=dtype)
self.register_buffer('dct_matrix', self.dct_matrix)
def forward(self, x):
log_mel_feature = self._log_melspectrogram(x)
mfcc = paddle.matmul(
log_mel_feature.transpose((0, 2, 1)), self.dct_matrix).transpose(
(0, 2, 1)) # (B, n_mels, L)
return mfcc

@ -0,0 +1,20 @@
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .functional import compute_fbank_matrix
from .functional import create_dct
from .functional import fft_frequencies
from .functional import hz_to_mel
from .functional import mel_frequencies
from .functional import mel_to_hz
from .functional import power_to_db

@ -0,0 +1,265 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Modified from librosa(https://github.com/librosa/librosa)
import math
from typing import Optional
from typing import Union
import paddle
__all__ = [
'hz_to_mel',
'mel_to_hz',
'mel_frequencies',
'fft_frequencies',
'compute_fbank_matrix',
'power_to_db',
'create_dct',
]
def hz_to_mel(freq: Union[paddle.Tensor, float],
htk: bool=False) -> Union[paddle.Tensor, float]:
"""Convert Hz to Mels.
Parameters:
freq: the input tensor of arbitrary shape, or a single floating point number.
htk: use HTK formula to do the conversion.
The default value is False.
Returns:
The frequencies represented in Mel-scale.
"""
if htk:
if isinstance(freq, paddle.Tensor):
return 2595.0 * paddle.log10(1.0 + freq / 700.0)
else:
return 2595.0 * math.log10(1.0 + freq / 700.0)
# Fill in the linear part
f_min = 0.0
f_sp = 200.0 / 3
mels = (freq - f_min) / f_sp
# Fill in the log-scale part
min_log_hz = 1000.0 # beginning of log region (Hz)
min_log_mel = (min_log_hz - f_min) / f_sp # same (Mels)
logstep = math.log(6.4) / 27.0 # step size for log region
if isinstance(freq, paddle.Tensor):
target = min_log_mel + paddle.log(
freq / min_log_hz + 1e-10) / logstep # prevent nan with 1e-10
mask = (freq > min_log_hz).astype(freq.dtype)
mels = target * mask + mels * (
1 - mask) # will replace by masked_fill OP in future
else:
if freq >= min_log_hz:
mels = min_log_mel + math.log(freq / min_log_hz + 1e-10) / logstep
return mels
def mel_to_hz(mel: Union[float, paddle.Tensor],
htk: bool=False) -> Union[float, paddle.Tensor]:
"""Convert mel bin numbers to frequencies.
Parameters:
mel: the mel frequency represented as a tensor of arbitrary shape, or a floating point number.
htk: use HTK formula to do the conversion.
Returns:
The frequencies represented in hz.
"""
if htk:
return 700.0 * (10.0**(mel / 2595.0) - 1.0)
f_min = 0.0
f_sp = 200.0 / 3
freqs = f_min + f_sp * mel
# And now the nonlinear scale
min_log_hz = 1000.0 # beginning of log region (Hz)
min_log_mel = (min_log_hz - f_min) / f_sp # same (Mels)
logstep = math.log(6.4) / 27.0 # step size for log region
if isinstance(mel, paddle.Tensor):
target = min_log_hz * paddle.exp(logstep * (mel - min_log_mel))
mask = (mel > min_log_mel).astype(mel.dtype)
freqs = target * mask + freqs * (
1 - mask) # will replace by masked_fill OP in future
else:
if mel >= min_log_mel:
freqs = min_log_hz * math.exp(logstep * (mel - min_log_mel))
return freqs
def mel_frequencies(n_mels: int=64,
f_min: float=0.0,
f_max: float=11025.0,
htk: bool=False,
dtype: str=paddle.float32):
"""Compute mel frequencies.
Parameters:
n_mels(int): number of Mel bins.
f_min(float): the lower cut-off frequency, below which the filter response is zero.
f_max(float): the upper cut-off frequency, above which the filter response is zero.
htk(bool): whether to use htk formula.
dtype(str): the datatype of the return frequencies.
Returns:
The frequencies represented in Mel-scale
"""
# 'Center freqs' of mel bands - uniformly spaced between limits
min_mel = hz_to_mel(f_min, htk=htk)
max_mel = hz_to_mel(f_max, htk=htk)
mels = paddle.linspace(min_mel, max_mel, n_mels, dtype=dtype)
freqs = mel_to_hz(mels, htk=htk)
return freqs
def fft_frequencies(sr: int, n_fft: int, dtype: str=paddle.float32):
"""Compute fourier frequencies.
Parameters:
sr(int): the audio sample rate.
n_fft(float): the number of fft bins.
dtype(str): the datatype of the return frequencies.
Returns:
The frequencies represented in hz.
"""
return paddle.linspace(0, float(sr) / 2, int(1 + n_fft // 2), dtype=dtype)
def compute_fbank_matrix(sr: int,
n_fft: int,
n_mels: int=64,
f_min: float=0.0,
f_max: Optional[float]=None,
htk: bool=False,
norm: Union[str, float]='slaney',
dtype: str=paddle.float32):
"""Compute fbank matrix.
Parameters:
sr(int): the audio sample rate.
n_fft(int): the number of fft bins.
n_mels(int): the number of Mel bins.
f_min(float): the lower cut-off frequency, below which the filter response is zero.
f_max(float): the upper cut-off frequency, above which the filter response is zero.
htk: whether to use htk formula.
return_complex(bool): whether to return complex matrix. If True, the matrix will
be complex type. Otherwise, the real and image part will be stored in the last
axis of returned tensor.
dtype(str): the datatype of the returned fbank matrix.
Returns:
The fbank matrix of shape (n_mels, int(1+n_fft//2)).
Shape:
output: (n_mels, int(1+n_fft//2))
"""
if f_max is None:
f_max = float(sr) / 2
# Initialize the weights
weights = paddle.zeros((n_mels, int(1 + n_fft // 2)), dtype=dtype)
# Center freqs of each FFT bin
fftfreqs = fft_frequencies(sr=sr, n_fft=n_fft, dtype=dtype)
# 'Center freqs' of mel bands - uniformly spaced between limits
mel_f = mel_frequencies(
n_mels + 2, f_min=f_min, f_max=f_max, htk=htk, dtype=dtype)
fdiff = mel_f[1:] - mel_f[:-1] #np.diff(mel_f)
ramps = mel_f.unsqueeze(1) - fftfreqs.unsqueeze(0)
#ramps = np.subtract.outer(mel_f, fftfreqs)
for i in range(n_mels):
# lower and upper slopes for all bins
lower = -ramps[i] / fdiff[i]
upper = ramps[i + 2] / fdiff[i + 1]
# .. then intersect them with each other and zero
weights[i] = paddle.maximum(
paddle.zeros_like(lower), paddle.minimum(lower, upper))
# Slaney-style mel is scaled to be approx constant energy per channel
if norm == 'slaney':
enorm = 2.0 / (mel_f[2:n_mels + 2] - mel_f[:n_mels])
weights *= enorm.unsqueeze(1)
elif isinstance(norm, int) or isinstance(norm, float):
weights = paddle.nn.functional.normalize(weights, p=norm, axis=-1)
return weights
def power_to_db(magnitude: paddle.Tensor,
ref_value: float=1.0,
amin: float=1e-10,
top_db: Optional[float]=None) -> paddle.Tensor:
"""Convert a power spectrogram (amplitude squared) to decibel (dB) units.
The function computes the scaling ``10 * log10(x / ref)`` in a numerically
stable way.
Parameters:
magnitude(Tensor): the input magnitude tensor of any shape.
ref_value(float): the reference value. If smaller than 1.0, the db level
of the signal will be pulled up accordingly. Otherwise, the db level
is pushed down.
amin(float): the minimum value of input magnitude, below which the input
magnitude is clipped(to amin).
top_db(float): the maximum db value of resulting spectrum, above which the
spectrum is clipped(to top_db).
Returns:
The spectrogram in log-scale.
shape:
input: any shape
output: same as input
"""
if amin <= 0:
raise Exception("amin must be strictly positive")
if ref_value <= 0:
raise Exception("ref_value must be strictly positive")
ones = paddle.ones_like(magnitude)
log_spec = 10.0 * paddle.log10(paddle.maximum(ones * amin, magnitude))
log_spec -= 10.0 * math.log10(max(ref_value, amin))
if top_db is not None:
if top_db < 0:
raise Exception("top_db must be non-negative")
log_spec = paddle.maximum(log_spec, ones * (log_spec.max() - top_db))
return log_spec
def create_dct(n_mfcc: int,
n_mels: int,
norm: Optional[str]='ortho',
dtype: Optional[str]=paddle.float32) -> paddle.Tensor:
"""Create a discrete cosine transform(DCT) matrix.
Parameters:
n_mfcc (int): Number of mel frequency cepstral coefficients.
n_mels (int): Number of mel filterbanks.
norm (str, optional): Normalizaiton type. Defaults to 'ortho'.
Returns:
Tensor: The DCT matrix with shape (n_mels, n_mfcc).
"""
n = paddle.arange(n_mels, dtype=dtype)
k = paddle.arange(n_mfcc, dtype=dtype).unsqueeze(1)
dct = paddle.cos(math.pi / float(n_mels) * (n + 0.5) *
k) # size (n_mfcc, n_mels)
if norm is None:
dct *= 2.0
else:
assert norm == "ortho"
dct[0] *= 1.0 / math.sqrt(2.0)
dct *= math.sqrt(2.0 / float(n_mels))
return dct.T

@ -20,6 +20,19 @@ from paddle import Tensor
__all__ = [
'get_window',
# windows
'taylor',
'hamming',
'hann',
'tukey',
'kaiser',
'gaussian',
'exponential',
'triang',
'bohman',
'blackman',
'cosine',
]
@ -73,6 +86,21 @@ def general_gaussian(M: int, p, sig, sym: bool=True,
return _truncate(w, needs_trunc)
def general_cosine(M: int, a: float, sym: bool=True,
dtype: str='float64') -> Tensor:
"""Compute a generic weighted sum of cosine terms window.
This function is consistent with scipy.signal.windows.general_cosine().
"""
if _len_guards(M):
return paddle.ones((M, ), dtype=dtype)
M, needs_trunc = _extend(M, sym)
fac = paddle.linspace(-math.pi, math.pi, M, dtype=dtype)
w = paddle.zeros((M, ), dtype=dtype)
for k in range(len(a)):
w += a[k] * paddle.cos(k * fac)
return _truncate(w, needs_trunc)
def general_hamming(M: int, alpha: float, sym: bool=True,
dtype: str='float64') -> Tensor:
"""Compute a generalized Hamming window.
@ -143,21 +171,6 @@ def taylor(M: int,
return _truncate(w, needs_trunc)
def general_cosine(M: int, a: float, sym: bool=True,
dtype: str='float64') -> Tensor:
"""Compute a generic weighted sum of cosine terms window.
This function is consistent with scipy.signal.windows.general_cosine().
"""
if _len_guards(M):
return paddle.ones((M, ), dtype=dtype)
M, needs_trunc = _extend(M, sym)
fac = paddle.linspace(-math.pi, math.pi, M, dtype=dtype)
w = paddle.zeros((M, ), dtype=dtype)
for k in range(len(a)):
w += a[k] * paddle.cos(k * fac)
return _truncate(w, needs_trunc)
def hamming(M: int, sym: bool=True, dtype: str='float64') -> Tensor:
"""Compute a Hamming window.
The Hamming window is a taper formed by using a raised cosine with
@ -375,6 +388,7 @@ def cosine(M: int, sym: bool=True, dtype: str='float64') -> Tensor:
return _truncate(w, needs_trunc)
## factory function
def get_window(window: Union[str, Tuple[str, float]],
win_length: int,
fftbins: bool=True,

@ -11,4 +11,3 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .audio import *

@ -1,6 +1,6 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
@ -11,8 +11,5 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .download import *
from .env import *
from .error import *
from .log import *
from .time import *
from .dtw import dtw_distance
from .mcd import mcd_distance

@ -0,0 +1,42 @@
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy as np
from dtaidistance import dtw_ndim
__all__ = [
'dtw_distance',
]
def dtw_distance(xs: np.ndarray, ys: np.ndarray) -> float:
"""dtw distance
Dynamic Time Warping.
This function keeps a compact matrix, not the full warping paths matrix.
Uses dynamic programming to compute:
wps[i, j] = (s1[i]-s2[j])**2 + min(
wps[i-1, j ] + penalty, // vertical / insertion / expansion
wps[i , j-1] + penalty, // horizontal / deletion / compression
wps[i-1, j-1]) // diagonal / match
dtw = sqrt(wps[-1, -1])
Args:
xs (np.ndarray): ref sequence, [T,D]
ys (np.ndarray): hyp sequence, [T,D]
Returns:
float: dtw distance
"""
return dtw_ndim.distance(xs, ys)

@ -0,0 +1,48 @@
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import mcd.metrics_fast as mt
import numpy as np
from mcd import dtw
__all__ = [
'mcd_distance',
]
def mcd_distance(xs: np.ndarray, ys: np.ndarray, cost_fn=mt.logSpecDbDist):
"""Mel cepstral distortion (MCD), dtw distance.
Dynamic Time Warping.
Uses dynamic programming to compute:
wps[i, j] = cost_fn(xs[i], ys[j]) + min(
wps[i-1, j ], // vertical / insertion / expansion
wps[i , j-1], // horizontal / deletion / compression
wps[i-1, j-1]) // diagonal / match
dtw = sqrt(wps[-1, -1])
Cost Function:
logSpecDbConst = 10.0 / math.log(10.0) * math.sqrt(2.0)
def logSpecDbDist(x, y):
diff = x - y
return logSpecDbConst * math.sqrt(np.inner(diff, diff))
Args:
xs (np.ndarray): ref sequence, [T,D]
ys (np.ndarray): hyp sequence, [T,D]
Returns:
float: dtw distance
"""
min_cost, path = dtw.dtw(xs, ys, cost_fn)
return min_cost

@ -0,0 +1,13 @@
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

@ -0,0 +1,25 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from .download import decompress
from .download import download_and_decompress
from .download import load_state_dict_from_url
from .env import DATA_HOME
from .env import MODEL_HOME
from .env import PPAUDIO_HOME
from .env import USER_HOME
from .error import ParameterError
from .log import Logger
from .log import logger
from .time import seconds_to_hms
from .time import Timer

@ -59,4 +59,4 @@ def load_state_dict_from_url(url: str, path: str, md5: str=None):
os.makedirs(path)
download.get_path_from_url(url, path, md5)
return load_state_dict(os.path.join(path, os.path.basename(url)))
return load_state_dict(os.path.join(path, os.path.basename(url)))

@ -20,6 +20,13 @@ PPAUDIO_HOME --> the root directory for storing PaddleAudio related data. D
'''
import os
__all__ = [
'USER_HOME',
'PPAUDIO_HOME',
'MODEL_HOME',
'DATA_HOME',
]
def _get_user_home():
return os.path.expanduser('~')

@ -19,7 +19,10 @@ import time
import colorlog
loggers = {}
__all__ = [
'Logger',
'logger',
]
log_config = {
'DEBUG': {

@ -14,6 +14,11 @@
import math
import time
__all__ = [
'Timer',
'seconds_to_hms',
]
class Timer(object):
'''Calculate runing speed and estimated time of arrival(ETA)'''

@ -14,7 +14,7 @@
import setuptools
# set the version here
VERSION = '0.1.0'
VERSION = '0.2.0'
def write_version_py(filename='paddleaudio/__init__.py'):
@ -59,6 +59,8 @@ setuptools.setup(
'resampy >= 0.2.2',
'soundfile >= 0.9.0',
'colorlog',
'dtaidistance >= 2.3.6',
'mcd >= 0.4',
], )
remove_version_py()

@ -11,3 +11,6 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import _locale
_locale._getdefaultlocale = (lambda *args: ['en_US', 'utf8'])

@ -34,7 +34,7 @@ def init(config):
bool:
"""
# init api
api_list = list(config.engine_backend)
api_list = list(engine.split("_")[0] for engine in config.engine_list)
api_router = setup_router(api_list)
app.include_router(api_router)

@ -62,7 +62,7 @@ class ServerExecutor(BaseExecutor):
bool:
"""
# init api
api_list = list(config.engine_backend)
api_list = list(engine.split("_")[0] for engine in config.engine_list)
api_router = setup_router(api_list)
app.include_router(api_router)

@ -1,27 +1,107 @@
# This is the parameter configuration file for PaddleSpeech Serving.
##################################################################
# SERVER SETTING #
##################################################################
host: '127.0.0.1'
#################################################################################
# SERVER SETTING #
#################################################################################
host: 127.0.0.1
port: 8090
##################################################################
# CONFIG FILE #
##################################################################
# add engine backend type (Options: asr, tts) and config file here.
# Adding a speech task to engine_backend means starting the service.
engine_backend:
asr: 'conf/asr/asr.yaml'
tts: 'conf/tts/tts.yaml'
# The engine_type of speech task needs to keep the same type as the config file of speech task.
# E.g: The engine_type of asr is 'python', the engine_backend of asr is 'XX/asr.yaml'
# E.g: The engine_type of asr is 'inference', the engine_backend of asr is 'XX/asr_pd.yaml'
#
# add engine type (Options: python, inference)
engine_type:
asr: 'python'
tts: 'python'
# The task format in the engin_list is: <speech task>_<engine type>
# task choices = ['asr_python', 'asr_inference', 'tts_python', 'tts_inference']
engine_list: ['asr_python', 'tts_python']
#################################################################################
# ENGINE CONFIG #
#################################################################################
################### speech task: asr; engine_type: python #######################
asr_python:
model: 'conformer_wenetspeech'
lang: 'zh'
sample_rate: 16000
cfg_path: # [optional]
ckpt_path: # [optional]
decode_method: 'attention_rescoring'
force_yes: True
device: # set 'gpu:id' or 'cpu'
################### speech task: asr; engine_type: inference #######################
asr_inference:
# model_type choices=['deepspeech2offline_aishell']
model_type: 'deepspeech2offline_aishell'
am_model: # the pdmodel file of am static model [optional]
am_params: # the pdiparams file of am static model [optional]
lang: 'zh'
sample_rate: 16000
cfg_path:
decode_method:
force_yes: True
am_predictor_conf:
device: # set 'gpu:id' or 'cpu'
switch_ir_optim: True
glog_info: False # True -> print glog
summary: True # False -> do not show predictor config
################### speech task: tts; engine_type: python #######################
tts_python:
# am (acoustic model) choices=['speedyspeech_csmsc', 'fastspeech2_csmsc',
# 'fastspeech2_ljspeech', 'fastspeech2_aishell3',
# 'fastspeech2_vctk']
am: 'fastspeech2_csmsc'
am_config:
am_ckpt:
am_stat:
phones_dict:
tones_dict:
speaker_dict:
spk_id: 0
# voc (vocoder) choices=['pwgan_csmsc', 'pwgan_ljspeech', 'pwgan_aishell3',
# 'pwgan_vctk', 'mb_melgan_csmsc']
voc: 'pwgan_csmsc'
voc_config:
voc_ckpt:
voc_stat:
# others
lang: 'zh'
device: # set 'gpu:id' or 'cpu'
################### speech task: tts; engine_type: inference #######################
tts_inference:
# am (acoustic model) choices=['speedyspeech_csmsc', 'fastspeech2_csmsc']
am: 'fastspeech2_csmsc'
am_model: # the pdmodel file of your am static model (XX.pdmodel)
am_params: # the pdiparams file of your am static model (XX.pdipparams)
am_sample_rate: 24000
phones_dict:
tones_dict:
speaker_dict:
spk_id: 0
am_predictor_conf:
device: # set 'gpu:id' or 'cpu'
switch_ir_optim: True
glog_info: False # True -> print glog
summary: True # False -> do not show predictor config
# voc (vocoder) choices=['pwgan_csmsc', 'mb_melgan_csmsc','hifigan_csmsc']
voc: 'pwgan_csmsc'
voc_model: # the pdmodel file of your vocoder static model (XX.pdmodel)
voc_params: # the pdiparams file of your vocoder static model (XX.pdipparams)
voc_sample_rate: 24000
voc_predictor_conf:
device: # set 'gpu:id' or 'cpu'
switch_ir_optim: True
glog_info: False # True -> print glog
summary: True # False -> do not show predictor config
# others
lang: 'zh'

@ -1,8 +0,0 @@
model: 'conformer_wenetspeech'
lang: 'zh'
sample_rate: 16000
cfg_path: # [optional]
ckpt_path: # [optional]
decode_method: 'attention_rescoring'
force_yes: True
device: # set 'gpu:id' or 'cpu'

@ -1,26 +0,0 @@
# This is the parameter configuration file for ASR server.
# These are the static models that support paddle inference.
##################################################################
# ACOUSTIC MODEL SETTING #
# am choices=['deepspeech2offline_aishell'] TODO
##################################################################
model_type: 'deepspeech2offline_aishell'
am_model: # the pdmodel file of am static model [optional]
am_params: # the pdiparams file of am static model [optional]
lang: 'zh'
sample_rate: 16000
cfg_path:
decode_method:
force_yes: True
am_predictor_conf:
device: # set 'gpu:id' or 'cpu'
switch_ir_optim: True
glog_info: False # True -> print glog
summary: True # False -> do not show predictor config
##################################################################
# OTHERS #
##################################################################

@ -1,32 +0,0 @@
# This is the parameter configuration file for TTS server.
##################################################################
# ACOUSTIC MODEL SETTING #
# am choices=['speedyspeech_csmsc', 'fastspeech2_csmsc',
# 'fastspeech2_ljspeech', 'fastspeech2_aishell3',
# 'fastspeech2_vctk']
##################################################################
am: 'fastspeech2_csmsc'
am_config:
am_ckpt:
am_stat:
phones_dict:
tones_dict:
speaker_dict:
spk_id: 0
##################################################################
# VOCODER SETTING #
# voc choices=['pwgan_csmsc', 'pwgan_ljspeech', 'pwgan_aishell3',
# 'pwgan_vctk', 'mb_melgan_csmsc']
##################################################################
voc: 'pwgan_csmsc'
voc_config:
voc_ckpt:
voc_stat:
##################################################################
# OTHERS #
##################################################################
lang: 'zh'
device: # set 'gpu:id' or 'cpu'

@ -1,42 +0,0 @@
# This is the parameter configuration file for TTS server.
# These are the static models that support paddle inference.
##################################################################
# ACOUSTIC MODEL SETTING #
# am choices=['speedyspeech_csmsc', 'fastspeech2_csmsc']
##################################################################
am: 'fastspeech2_csmsc'
am_model: # the pdmodel file of your am static model (XX.pdmodel)
am_params: # the pdiparams file of your am static model (XX.pdipparams)
am_sample_rate: 24000
phones_dict:
tones_dict:
speaker_dict:
spk_id: 0
am_predictor_conf:
device: # set 'gpu:id' or 'cpu'
switch_ir_optim: True
glog_info: False # True -> print glog
summary: True # False -> do not show predictor config
##################################################################
# VOCODER SETTING #
# voc choices=['pwgan_csmsc', 'mb_melgan_csmsc','hifigan_csmsc']
##################################################################
voc: 'pwgan_csmsc'
voc_model: # the pdmodel file of your vocoder static model (XX.pdmodel)
voc_params: # the pdiparams file of your vocoder static model (XX.pdipparams)
voc_sample_rate: 24000
voc_predictor_conf:
device: # set 'gpu:id' or 'cpu'
switch_ir_optim: True
glog_info: False # True -> print glog
summary: True # False -> do not show predictor config
##################################################################
# OTHERS #
##################################################################
lang: 'zh'

@ -26,7 +26,6 @@ from paddlespeech.s2t.frontend.featurizer.text_featurizer import TextFeaturizer
from paddlespeech.s2t.modules.ctc import CTCDecoder
from paddlespeech.s2t.utils.utility import UpdateConfig
from paddlespeech.server.engine.base_engine import BaseEngine
from paddlespeech.server.utils.config import get_config
from paddlespeech.server.utils.paddle_predictor import init_predictor
from paddlespeech.server.utils.paddle_predictor import run_model
@ -184,7 +183,7 @@ class ASREngine(BaseEngine):
def __init__(self):
super(ASREngine, self).__init__()
def init(self, config_file: str) -> bool:
def init(self, config: dict) -> bool:
"""init engine resource
Args:
@ -196,7 +195,7 @@ class ASREngine(BaseEngine):
self.input = None
self.output = None
self.executor = ASRServerExecutor()
self.config = get_config(config_file)
self.config = config
self.executor._init_from_path(
model_type=self.config.model_type,

@ -19,7 +19,6 @@ import paddle
from paddlespeech.cli.asr.infer import ASRExecutor
from paddlespeech.cli.log import logger
from paddlespeech.server.engine.base_engine import BaseEngine
from paddlespeech.server.utils.config import get_config
__all__ = ['ASREngine']
@ -40,7 +39,7 @@ class ASREngine(BaseEngine):
def __init__(self):
super(ASREngine, self).__init__()
def init(self, config_file: str) -> bool:
def init(self, config: dict) -> bool:
"""init engine resource
Args:
@ -52,8 +51,7 @@ class ASREngine(BaseEngine):
self.input = None
self.output = None
self.executor = ASRServerExecutor()
self.config = get_config(config_file)
self.config = config
try:
if self.config.device:
self.device = self.config.device

@ -28,11 +28,13 @@ def init_engine_pool(config) -> bool:
""" Init engine pool
"""
global ENGINE_POOL
for engine in config.engine_backend:
for engine_and_type in config.engine_list:
engine = engine_and_type.split("_")[0]
engine_type = engine_and_type.split("_")[1]
ENGINE_POOL[engine] = EngineFactory.get_engine(
engine_name=engine, engine_type=config.engine_type[engine])
if not ENGINE_POOL[engine].init(
config_file=config.engine_backend[engine]):
engine_name=engine, engine_type=engine_type)
if not ENGINE_POOL[engine].init(config=config[engine_and_type]):
return False
return True

@ -29,7 +29,6 @@ from paddlespeech.cli.utils import download_and_decompress
from paddlespeech.cli.utils import MODEL_HOME
from paddlespeech.server.engine.base_engine import BaseEngine
from paddlespeech.server.utils.audio_process import change_speed
from paddlespeech.server.utils.config import get_config
from paddlespeech.server.utils.errors import ErrorCode
from paddlespeech.server.utils.exception import ServerBaseException
from paddlespeech.server.utils.paddle_predictor import init_predictor
@ -357,11 +356,11 @@ class TTSEngine(BaseEngine):
"""
super(TTSEngine, self).__init__()
def init(self, config_file: str) -> bool:
def init(self, config: dict) -> bool:
self.executor = TTSServerExecutor()
try:
self.config = get_config(config_file)
self.config = config
self.executor._init_from_path(
am=self.config.am,
am_model=self.config.am_model,

@ -25,7 +25,6 @@ from paddlespeech.cli.log import logger
from paddlespeech.cli.tts.infer import TTSExecutor
from paddlespeech.server.engine.base_engine import BaseEngine
from paddlespeech.server.utils.audio_process import change_speed
from paddlespeech.server.utils.config import get_config
from paddlespeech.server.utils.errors import ErrorCode
from paddlespeech.server.utils.exception import ServerBaseException
@ -50,11 +49,11 @@ class TTSEngine(BaseEngine):
"""
super(TTSEngine, self).__init__()
def init(self, config_file: str) -> bool:
def init(self, config: dict) -> bool:
self.executor = TTSServerExecutor()
try:
self.config = get_config(config_file)
self.config = config
if self.config.device:
self.device = self.config.device
else:

@ -34,7 +34,7 @@ def main():
"--generator-type",
type=str,
default="pwgan",
help="type of GANVocoder, should in {pwgan, mb_melgan, style_melgan, } now"
help="type of GANVocoder, should in {pwgan, mb_melgan, style_melgan, hifigan, } now"
)
parser.add_argument("--config", type=str, help="GANVocoder config file.")
parser.add_argument("--checkpoint", type=str, help="snapshot to load.")

@ -11,3 +11,31 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
__init__ file for sidt package.
"""
import logging as sidt_logging
import colorlog
LOG_COLOR_CONFIG = {
'DEBUG': 'white',
'INFO': 'white',
'WARNING': 'yellow',
'ERROR': 'red',
'CRITICAL': 'purple',
}
# 设置全局的logger
colored_formatter = colorlog.ColoredFormatter(
'%(log_color)s [%(levelname)s] [%(asctime)s] [%(filename)s:%(lineno)d] - %(message)s',
datefmt="%Y-%m-%d %H:%M:%S",
log_colors=LOG_COLOR_CONFIG) # 日志输出格式
_logger = sidt_logging.getLogger("sidt")
handler = colorlog.StreamHandler()
handler.setLevel(sidt_logging.INFO)
handler.setFormatter(colored_formatter)
_logger.addHandler(handler)
_logger.setLevel(sidt_logging.INFO)

@ -0,0 +1,142 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import sys
import random
import numpy as np
import kaldi_python_io as k_io
from paddle.io import Dataset
from paddlespeech.vector.utils.data_utils import batch_pad_right
import paddlespeech.vector.utils as utils
from paddlespeech.vector.utils.utils import read_map_file
from paddlespeech.vector import _logger as log
def ark_collate_fn(batch):
"""
Custom collate function] for kaldi feats dataset
Args:
min_chunk_size: min chunk size of a utterance
max_chunk_size: max chunk size of a utterance
Returns:
ark_collate_fn: collate funtion for dataloader
"""
data = []
target = []
for items in batch:
for x, y in zip(items[0], items[1]):
data.append(np.array(x))
target.append(y)
data, lengths = batch_pad_right(data)
return np.array(data, dtype=np.float32), \
np.array(lengths, dtype=np.float32), \
np.array(target, dtype=np.long).reshape((len(target), 1))
class KaldiArkDataset(Dataset):
"""
Dataset used to load kaldi ark/scp files.
"""
def __init__(self, scp_file, label2utt, min_item_size=1,
max_item_size=1, repeat=50, min_chunk_size=200,
max_chunk_size=400, select_by_speaker=True):
self.scp_file = scp_file
self.scp_reader = None
self.repeat = repeat
self.min_item_size = min_item_size
self.max_item_size = max_item_size
self.min_chunk_size = min_chunk_size
self.max_chunk_size = max_chunk_size
self._collate_fn = ark_collate_fn
self._is_select_by_speaker = select_by_speaker
if utils.is_exist(self.scp_file):
self.scp_reader = k_io.ScriptReader(self.scp_file)
label2utts, utt2label = read_map_file(label2utt, key_func=int)
self.utt_info = list(label2utts.items()) if self._is_select_by_speaker else list(utt2label.items())
@property
def collate_fn(self):
"""
Return a collate funtion.
"""
return self._collate_fn
def _random_chunk(self, length):
chunk_size = random.randint(self.min_chunk_size, self.max_chunk_size)
if chunk_size >= length:
return 0, length
start = random.randint(0, length - chunk_size)
end = start + chunk_size
return start, end
def _select_by_speaker(self, index):
if self.scp_reader is None or not self.utt_info:
return []
index = index % (len(self.utt_info))
inputs = []
labels = []
item_size = random.randint(self.min_item_size, self.max_item_size)
for loop_idx in range(item_size):
try:
utt_index = random.randint(0, len(self.utt_info[index][1])) \
% len(self.utt_info[index][1])
key = self.utt_info[index][1][utt_index]
except:
print(index, utt_index, len(self.utt_info[index][1]))
sys.exit(-1)
x = self.scp_reader[key]
x = np.transpose(x)
bg, end = self._random_chunk(x.shape[-1])
inputs.append(x[:, bg: end])
labels.append(self.utt_info[index][0])
return inputs, labels
def _select_by_utt(self, index):
if self.scp_reader is None or len(self.utt_info) == 0:
return {}
index = index % (len(self.utt_info))
key = self.utt_info[index][0]
x = self.scp_reader[key]
x = np.transpose(x)
bg, end = self._random_chunk(x.shape[-1])
y = self.utt_info[index][1]
return [x[:, bg: end]], [y]
def __getitem__(self, index):
if self._is_select_by_speaker:
return self._select_by_speaker(index)
else:
return self._select_by_utt(index)
def __len__(self):
return len(self.utt_info) * self.repeat
def __iter__(self):
self._start = 0
return self
def __next__(self):
if self._start < len(self):
ret = self[self._start]
self._start += 1
return ret
else:
raise StopIteration

@ -0,0 +1,143 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import sys
import random
import numpy as np
import kaldi_python_io as k_io
from paddle.io import Dataset
from paddlespeech.vector.utils.data_utils import batch_pad_right
import paddlespeech.vector.utils as utils
from paddlespeech.vector.utils.utils import read_map_file
def ark_collate_fn(batch):
"""
Custom collate function for kaldi feats dataset
Args:
min_chunk_size: min chunk size of a utterance
max_chunk_size: max chunk size of a utterance
Returns:
ark_collate_fn: collate funtion for dataloader
"""
data = []
target = []
for items in batch:
for x, y in zip(items[0], items[1]):
data.append(np.array(x))
target.append(y)
data, lengths = batch_pad_right(data)
return np.array(data, dtype=np.float32), \
np.array(lengths, dtype=np.float32), \
np.array(target, dtype=np.long).reshape((len(target), 1))
class KaldiArkDataset(Dataset):
"""
Dataset used to load kaldi ark/scp files.
"""
def __init__(self, scp_file, label2utt, min_item_size=1,
max_item_size=1, repeat=50, min_chunk_size=200,
max_chunk_size=400, select_by_speaker=True):
self.scp_file = scp_file
self.scp_reader = None
self.repeat = repeat
self.min_item_size = min_item_size
self.max_item_size = max_item_size
self.min_chunk_size = min_chunk_size
self.max_chunk_size = max_chunk_size
self._collate_fn = ark_collate_fn
self._is_select_by_speaker = select_by_speaker
if utils.is_exist(self.scp_file):
self.scp_reader = k_io.ScriptReader(self.scp_file)
label2utts, utt2label = read_map_file(label2utt, key_func=int)
self.utt_info = list(label2utts.items()) if self._is_select_by_speaker else list(utt2label.items())
@property
def collate_fn(self):
"""
Return a collate funtion.
"""
return self._collate_fn
def _random_chunk(self, length):
chunk_size = random.randint(self.min_chunk_size, self.max_chunk_size)
if chunk_size >= length:
return 0, length
start = random.randint(0, length - chunk_size)
end = start + chunk_size
return start, end
def _select_by_speaker(self, index):
if self.scp_reader is None or not self.utt_info:
return []
index = index % (len(self.utt_info))
inputs = []
labels = []
item_size = random.randint(self.min_item_size, self.max_item_size)
for loop_idx in range(item_size):
try:
utt_index = random.randint(0, len(self.utt_info[index][1])) \
% len(self.utt_info[index][1])
key = self.utt_info[index][1][utt_index]
except:
print(index, utt_index, len(self.utt_info[index][1]))
sys.exit(-1)
x = self.scp_reader[key]
x = np.transpose(x)
bg, end = self._random_chunk(x.shape[-1])
inputs.append(x[:, bg: end])
labels.append(self.utt_info[index][0])
return inputs, labels
def _select_by_utt(self, index):
if self.scp_reader is None or len(self.utt_info) == 0:
return {}
index = index % (len(self.utt_info))
key = self.utt_info[index][0]
x = self.scp_reader[key]
x = np.transpose(x)
bg, end = self._random_chunk(x.shape[-1])
y = self.utt_info[index][1]
return [x[:, bg: end]], [y]
def __getitem__(self, index):
if self._is_select_by_speaker:
return self._select_by_speaker(index)
else:
return self._select_by_utt(index)
def __len__(self):
return len(self.utt_info) * self.repeat
def __iter__(self):
self._start = 0
return self
def __next__(self):
if self._start < len(self):
ret = self[self._start]
self._start += 1
return ret
else:
raise StopIteration
return KaldiArkDataset

@ -0,0 +1,91 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Load nnet3 training egs which generated by kaldi
"""
import random
import numpy as np
import kaldi_python_io as k_io
from paddle.io import Dataset
import paddlespeech.vector.utils.utils as utils
from paddlespeech.vector import _logger as log
class KaldiEgsDataset(Dataset):
"""
Dataset used to load kaldi nnet3 egs files.
"""
def __init__(self, egs_list_file, egs_idx, transforms=None):
self.scp_reader = None
self.subset_idx = egs_idx - 1
self.transforms = transforms
if not utils.is_exist(egs_list_file):
return
self.egs_files = []
with open(egs_list_file, 'r') as in_fh:
for line in in_fh:
if line.strip():
self.egs_files.append(line.strip())
self.next_subset()
def next_subset(self, target_index=None, delta_index=None):
"""
Use next specific subset
Args:
target_index: target egs index
delta_index: incremental value of egs index
"""
if self.egs_files:
if target_index:
self.subset_idx = target_index
else:
delta_index = delta_index if delta_index else 1
self.subset_idx += delta_index
log.info("egs dataset subset index: %d" % (self.subset_idx))
egs_file = self.egs_files[self.subset_idx % len(self.egs_files)]
if utils.is_exist(egs_file):
self.scp_reader = k_io.Nnet3EgsScriptReader(egs_file)
else:
log.warning("No such file or directory: %s" % (egs_file))
def __getitem__(self, index):
if self.scp_reader is None:
return {}
index %= len(self)
in_dict, out_dict = self.scp_reader[index]
x = np.array(in_dict['matrix'])
x = np.transpose(x)
y = np.array(out_dict['matrix'][0][0][0], dtype=np.int).reshape((1,))
if self.transforms is not None:
idx = random.randint(0, len(self.transforms) - 1)
x = self.transforms[idx](x)
return x, y
def __len__(self):
return len(self.scp_reader)
def __iter__(self):
self._start = 0
return self
def __next__(self):
if self._start < len(self):
ret = self[self._start]
self._start += 1
return ret
else:
raise StopIteration

@ -0,0 +1,125 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
data utilities
"""
import os
import sys
import numpy
import paddle
def pad_right_to(array, target_shape, mode="constant", value=0):
"""
This function takes a numpy array of arbitrary shape and pads it to target
shape by appending values on the right.
Args:
array: input numpy array. Input array whose dimension we need to pad.
target_shape : (list, tuple). Target shape we want for the target array its len must be equal to array.ndim
mode : str. Pad mode, please refer to numpy.pad documentation.
value : float. Pad value, please refer to numpy.pad documentation.
Returns:
array: numpy.array. Padded array.
valid_vals : list. List containing proportion for each dimension of original, non-padded values.
"""
assert len(target_shape) == array.ndim
pads = [] # this contains the abs length of the padding for each dimension.
valid_vals = [] # thic contains the relative lengths for each dimension.
i = 0 # iterating over target_shape ndims
while i < len(target_shape):
assert (
target_shape[i] >= array.shape[i]
), "Target shape must be >= original shape for every dim"
pads.append([0, target_shape[i] - array.shape[i]])
valid_vals.append(array.shape[i] / target_shape[i])
i += 1
array = numpy.pad(array, pads, mode=mode, constant_values=value)
return array, valid_vals
def batch_pad_right(arrays, mode="constant", value=0):
"""Given a list of numpy arrays it batches them together by padding to the right
on each dimension in order to get same length for all.
Args:
arrays : list. List of array we wish to pad together.
mode : str. Padding mode see numpy.pad documentation.
value : float. Padding value see numpy.pad documentation.
Returns:
array : numpy.array. Padded array.
valid_vals : list. List containing proportion for each dimension of original, non-padded values.
"""
if not len(arrays):
raise IndexError("arrays list must not be empty")
if len(arrays) == 1:
# if there is only one array in the batch we simply unsqueeze it.
return numpy.expand_dims(arrays[0], axis=0), numpy.array([1.0])
if not (
any(
[arrays[i].ndim == arrays[0].ndim for i in range(1, len(arrays))]
)
):
raise IndexError("All arrays must have same number of dimensions")
# FIXME we limit the support here: we allow padding of only the last dimension
# need to remove this when feat extraction is updated to handle multichannel.
max_shape = []
for dim in range(arrays[0].ndim):
if dim != (arrays[0].ndim - 1):
if not all(
[x.shape[dim] == arrays[0].shape[dim] for x in arrays[1:]]
):
raise EnvironmentError(
"arrays should have same dimensions except for last one"
)
max_shape.append(max([x.shape[dim] for x in arrays]))
batched = []
valid = []
for t in arrays:
# for each array we apply pad_right_to
padded, valid_percent = pad_right_to(
t, max_shape, mode=mode, value=value
)
batched.append(padded)
valid.append(valid_percent[-1])
batched = numpy.stack(batched)
return batched, numpy.array(valid)
def length_to_mask(length, max_len=None, dtype=None):
"""Creates a binary mask for each sequence.
"""
assert len(length.shape) == 1
if max_len is None:
max_len = paddle.cast(paddle.max(length), dtype="int64") # using arange to generate mask
mask = paddle.arange(max_len, dtype=length.dtype).expand([paddle.shape(length)[0], max_len]) < length.unsqueeze(1)
if dtype is None:
dtype = length.dtype
mask = paddle.cast(mask, dtype=dtype)
return mask

@ -0,0 +1,132 @@
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
utilities
"""
import os
import sys
import paddle
import numpy as np
from paddlespeech.vector import _logger as log
def exit_if_not_exist(in_path):
"""
Check the existence of a file or directory, if not exit, exit the program.
Args:
in_path: input dicrector
"""
if not is_exist(in_path):
sys.exit(-1)
def is_exist(in_path):
"""
Check the existence of a file or directory
Args:
in_path: input dicrector
Returns:
True or False
"""
if not os.path.exists(in_path):
log.error("No such file or directory: %s" % (in_path))
return False
return True
def get_latest_file(target_dir):
"""
Get the latest file in target directory
Args:
target_dir: target directory
Returns:
latest_file: a string or None
"""
items = os.listdir(target_dir)
items.sort(key=lambda fn: os.path.getmtime(os.path.join(target_dir, fn)) \
if not os.path.isdir(os.path.join(target_dir, fn)) else 0)
latest_file = None if not items else os.path.join(target_dir, items[-1])
return latest_file
def avg_models(models):
"""
merge multiple models
"""
checkpoint_dict = paddle.load(models[0])
final_state_dict = checkpoint_dict
if len(models) > 1:
for model in models[1:]:
checkpoint_dict = paddle.load(model)
for k, v in checkpoint_dict.items():
final_state_dict[k] += v
for k in final_state_dict.keys():
final_state_dict[k] /= float(len(models))
if np.any(np.isnan(final_state_dict[k])):
print("Nan in %s" % (k))
return final_state_dict
def Q_from_tokens(token_num):
"""
get prior model, data from uniform, would support others(guassian) in future
"""
freq = [1] * token_num
Q = paddle.to_tensor(freq, dtype = 'float64')
return Q / Q.sum()
def read_map_file(map_file, key_func=None, value_func=None, values_func=None):
""" Read map file. First colume is key, the rest columes are values.
Args:
map_file: map file
key_func: convert function for key
value_func: convert function for each value
values_func: convert function for values
Returns:
dict: key 2 value
dict: value 2 key
"""
if not is_exist(map_file):
sys.exit(0)
key2val = {}
val2key = {}
with open(map_file, 'r') as f:
for line in f:
line = line.strip()
if not line:
continue
items = line.split()
assert len(items) >= 2
key = items[0] if not key_func else key_func(items[0])
values = items[1:] if not value_func else [value_func(item) for item in items[1:]]
if values_func:
values = values_func(values)
key2val[key] = values
for value in values:
val2key[value] = key
return key2val, val2key

@ -0,0 +1,105 @@
#!/usr/bin/python
import argparse
import os
import yaml
def change_device(yamlfile: str, engine: str, device: str):
"""Change the settings of the device under the voice task configuration file
Args:
yaml_name (str): asr or asr_pd or tts or tts_pd
cpu (bool): True means set device to "cpu"
model_type (dict): change model type
"""
tmp_yamlfile = yamlfile.split(".yaml")[0] + "_tmp.yaml"
os.system("cp %s %s" % (yamlfile, tmp_yamlfile))
if device == 'cpu':
set_device = 'cpu'
elif device == 'gpu':
set_device = 'gpu:0'
else:
print("Please set correct device: cpu or gpu.")
with open(tmp_yamlfile) as f, open(yamlfile, "w+", encoding="utf-8") as fw:
y = yaml.safe_load(f)
if engine == 'asr_python' or engine == 'tts_python':
y[engine]['device'] = set_device
elif engine == 'asr_inference':
y[engine]['am_predictor_conf']['device'] = set_device
elif engine == 'tts_inference':
y[engine]['am_predictor_conf']['device'] = set_device
y[engine]['voc_predictor_conf']['device'] = set_device
else:
print(
"Please set correct engine: asr_python, tts_python, asr_inference, tts_inference."
)
print(yaml.dump(y, default_flow_style=False, sort_keys=False))
yaml.dump(y, fw, allow_unicode=True)
os.system("rm %s" % (tmp_yamlfile))
print("Change %s successfully." % (yamlfile))
def change_engine_type(yamlfile: str, engine_type):
"""Change the engine type and corresponding configuration file of the speech task in application.yaml
Args:
task (str): asr or tts
"""
tmp_yamlfile = yamlfile.split(".yaml")[0] + "_tmp.yaml"
os.system("cp %s %s" % (yamlfile, tmp_yamlfile))
speech_task = engine_type.split("_")[0]
with open(tmp_yamlfile) as f, open(yamlfile, "w+", encoding="utf-8") as fw:
y = yaml.safe_load(f)
engine_list = y['engine_list']
for engine in engine_list:
if speech_task in engine:
engine_list.remove(engine)
engine_list.append(engine_type)
y['engine_list'] = engine_list
print(yaml.dump(y, default_flow_style=False, sort_keys=False))
yaml.dump(y, fw, allow_unicode=True)
os.system("rm %s" % (tmp_yamlfile))
print("Change %s successfully." % (yamlfile))
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument(
'--config_file',
type=str,
default='./conf/application.yaml',
help='server yaml file.')
parser.add_argument(
'--change_task',
type=str,
default=None,
help='Change task',
choices=[
'enginetype-asr_python',
'enginetype-asr_inference',
'enginetype-tts_python',
'enginetype-tts_inference',
'device-asr_python-cpu',
'device-asr_python-gpu',
'device-asr_inference-cpu',
'device-asr_inference-gpu',
'device-tts_python-cpu',
'device-tts_python-gpu',
'device-tts_inference-cpu',
'device-tts_inference-gpu',
],
required=True)
args = parser.parse_args()
types = args.change_task.split("-")
if types[0] == "enginetype":
change_engine_type(args.config_file, types[1])
elif types[0] == "device":
change_device(args.config_file, types[1], types[2])
else:
print("Error change task, please check change_task.")

@ -0,0 +1,107 @@
# This is the parameter configuration file for PaddleSpeech Serving.
#################################################################################
# SERVER SETTING #
#################################################################################
host: 127.0.0.1
port: 8090
# The task format in the engin_list is: <speech task>_<engine type>
# task choices = ['asr_python', 'asr_inference', 'tts_python', 'tts_inference']
engine_list: ['asr_python', 'tts_python']
#################################################################################
# ENGINE CONFIG #
#################################################################################
################### speech task: asr; engine_type: python #######################
asr_python:
model: 'conformer_wenetspeech'
lang: 'zh'
sample_rate: 16000
cfg_path: # [optional]
ckpt_path: # [optional]
decode_method: 'attention_rescoring'
force_yes: True
device: # set 'gpu:id' or 'cpu'
################### speech task: asr; engine_type: inference #######################
asr_inference:
# model_type choices=['deepspeech2offline_aishell']
model_type: 'deepspeech2offline_aishell'
am_model: # the pdmodel file of am static model [optional]
am_params: # the pdiparams file of am static model [optional]
lang: 'zh'
sample_rate: 16000
cfg_path:
decode_method:
force_yes: True
am_predictor_conf:
device: # set 'gpu:id' or 'cpu'
switch_ir_optim: True
glog_info: False # True -> print glog
summary: True # False -> do not show predictor config
################### speech task: tts; engine_type: python #######################
tts_python:
# am (acoustic model) choices=['speedyspeech_csmsc', 'fastspeech2_csmsc',
# 'fastspeech2_ljspeech', 'fastspeech2_aishell3',
# 'fastspeech2_vctk']
am: 'fastspeech2_csmsc'
am_config:
am_ckpt:
am_stat:
phones_dict:
tones_dict:
speaker_dict:
spk_id: 0
# voc (vocoder) choices=['pwgan_csmsc', 'pwgan_ljspeech', 'pwgan_aishell3',
# 'pwgan_vctk', 'mb_melgan_csmsc']
voc: 'pwgan_csmsc'
voc_config:
voc_ckpt:
voc_stat:
# others
lang: 'zh'
device: # set 'gpu:id' or 'cpu'
################### speech task: tts; engine_type: inference #######################
tts_inference:
# am (acoustic model) choices=['speedyspeech_csmsc', 'fastspeech2_csmsc']
am: 'fastspeech2_csmsc'
am_model: # the pdmodel file of your am static model (XX.pdmodel)
am_params: # the pdiparams file of your am static model (XX.pdipparams)
am_sample_rate: 24000
phones_dict:
tones_dict:
speaker_dict:
spk_id: 0
am_predictor_conf:
device: # set 'gpu:id' or 'cpu'
switch_ir_optim: True
glog_info: False # True -> print glog
summary: True # False -> do not show predictor config
# voc (vocoder) choices=['pwgan_csmsc', 'mb_melgan_csmsc','hifigan_csmsc']
voc: 'pwgan_csmsc'
voc_model: # the pdmodel file of your vocoder static model (XX.pdmodel)
voc_params: # the pdiparams file of your vocoder static model (XX.pdipparams)
voc_sample_rate: 24000
voc_predictor_conf:
device: # set 'gpu:id' or 'cpu'
switch_ir_optim: True
glog_info: False # True -> print glog
summary: True # False -> do not show predictor config
# others
lang: 'zh'

@ -0,0 +1,186 @@
#!/bin/bash
# bash test_server_client.sh
StartService(){
# Start service
paddlespeech_server start --config_file $config_file 1>>log/server.log 2>>log/server.log.wf &
echo $! > pid
start_num=$(cat log/server.log.wf | grep "INFO: Uvicorn running on http://" -c)
flag="normal"
while [[ $start_num -lt $target_start_num && $flag == "normal" ]]
do
start_num=$(cat log/server.log.wf | grep "INFO: Uvicorn running on http://" -c)
# start service failed
if [ $(cat log/server.log.wf | grep -i "error" -c) -gt $error_time ];then
echo "Service started failed." | tee -a ./log/test_result.log
error_time=$(cat log/server.log.wf | grep -i "error" -c)
flag="unnormal"
fi
done
}
ClientTest(){
# Client test
# test asr client
paddlespeech_client asr --server_ip $server_ip --port $port --input ./zh.wav
((test_times+=1))
paddlespeech_client asr --server_ip $server_ip --port $port --input ./zh.wav
((test_times+=1))
# test tts client
paddlespeech_client tts --server_ip $server_ip --port $port --input "您好,欢迎使用百度飞桨语音合成服务。" --output output.wav
((test_times+=1))
paddlespeech_client tts --server_ip $server_ip --port $port --input "您好,欢迎使用百度飞桨语音合成服务。" --output output.wav
((test_times+=1))
}
GetTestResult() {
# Determine if the test was successful
response_success_time=$(cat log/server.log | grep "200 OK" -c)
if (( $response_success_time == $test_times )) ; then
echo "Testing successfully. The service configuration is: asr engine type: $1; tts engine type: $1; device: $2." | tee -a ./log/test_result.log
else
echo "Testing failed. The service configuration is: asr engine type: $1; tts engine type: $1; device: $2." | tee -a ./log/test_result.log
fi
test_times=$response_success_time
}
mkdir -p log
rm -rf log/server.log.wf
rm -rf log/server.log
rm -rf log/test_result.log
config_file=./conf/application.yaml
server_ip=$(cat $config_file | grep "host" | awk -F " " '{print $2}')
port=$(cat $config_file | grep "port" | awk '/port:/ {print $2}')
echo "Sevice ip: $server_ip" | tee ./log/test_result.log
echo "Sevice port: $port" | tee -a ./log/test_result.log
# whether a process is listening on $port
pid=`lsof -i :"$port"|grep -v "PID" | awk '{print $2}'`
if [ "$pid" != "" ]; then
echo "The port: $port is occupied, please change another port"
exit
fi
# download test audios for ASR client
wget -c https://paddlespeech.bj.bcebos.com/PaddleAudio/zh.wav https://paddlespeech.bj.bcebos.com/PaddleAudio/en.wav
target_start_num=0 # the number of start service
test_times=0 # The number of client test
error_time=0 # The number of error occurrences in the startup failure server.log.wf file
# start server: asr engine type: python; tts engine type: python; device: gpu
echo "Start the service: asr engine type: python; tts engine type: python; device: gpu" | tee -a ./log/test_result.log
((target_start_num+=1))
StartService
if [[ $start_num -eq $target_start_num && $flag == "normal" ]]; then
echo "Service started successfully." | tee -a ./log/test_result.log
ClientTest
echo "This round of testing is over." | tee -a ./log/test_result.log
GetTestResult python gpu
else
echo "Service failed to start, no client test."
target_start_num=$start_num
fi
kill -9 `cat pid`
rm -rf pid
sleep 2s
echo "**************************************************************************************" | tee -a ./log/test_result.log
# start server: asr engine type: python; tts engine type: python; device: cpu
python change_yaml.py --change_task device-asr_python-cpu # change asr.yaml device: cpu
python change_yaml.py --change_task device-tts_python-cpu # change tts.yaml device: cpu
echo "Start the service: asr engine type: python; tts engine type: python; device: cpu" | tee -a ./log/test_result.log
((target_start_num+=1))
StartService
if [[ $start_num -eq $target_start_num && $flag == "normal" ]]; then
echo "Service started successfully." | tee -a ./log/test_result.log
ClientTest
echo "This round of testing is over." | tee -a ./log/test_result.log
GetTestResult python cpu
else
echo "Service failed to start, no client test."
target_start_num=$start_num
fi
kill -9 `cat pid`
rm -rf pid
sleep 2s
echo "**************************************************************************************" | tee -a ./log/test_result.log
# start server: asr engine type: inference; tts engine type: inference; device: gpu
python change_yaml.py --change_task enginetype-asr_inference # change application.yaml, asr engine_type: inference; asr engine_backend: asr_pd.yaml
python change_yaml.py --change_task enginetype-tts_inference # change application.yaml, tts engine_type: inference; tts engine_backend: tts_pd.yaml
echo "Start the service: asr engine type: inference; tts engine type: inference; device: gpu" | tee -a ./log/test_result.log
((target_start_num+=1))
StartService
if [[ $start_num -eq $target_start_num && $flag == "normal" ]]; then
echo "Service started successfully." | tee -a ./log/test_result.log
ClientTest
echo "This round of testing is over." | tee -a ./log/test_result.log
GetTestResult inference gpu
else
echo "Service failed to start, no client test."
target_start_num=$start_num
fi
kill -9 `cat pid`
rm -rf pid
sleep 2s
echo "**************************************************************************************" | tee -a ./log/test_result.log
# start server: asr engine type: inference; tts engine type: inference; device: cpu
python change_yaml.py --change_task device-asr_inference-cpu # change asr_pd.yaml device: cpu
python change_yaml.py --change_task device-tts_inference-cpu # change tts_pd.yaml device: cpu
echo "start the service: asr engine type: inference; tts engine type: inference; device: cpu" | tee -a ./log/test_result.log
((target_start_num+=1))
StartService
if [[ $start_num -eq $target_start_num && $flag == "normal" ]]; then
echo "Service started successfully." | tee -a ./log/test_result.log
ClientTest
echo "This round of testing is over." | tee -a ./log/test_result.log
GetTestResult inference cpu
else
echo "Service failed to start, no client test."
target_start_num=$start_num
fi
kill -9 `cat pid`
rm -rf pid
sleep 2s
echo "**************************************************************************************" | tee -a ./log/test_result.log
echo "All tests completed." | tee -a ./log/test_result.log
# sohw all the test results
echo "***************** Here are all the test results ********************"
cat ./log/test_result.log
# Restoring conf is the same as demos/speech_server
rm -rf ./conf
cp ../../../demos/speech_server/conf/ ./ -rf
Loading…
Cancel
Save