asr-deepspeech 0.3.2

ASRDeepspeech (English / Japanese)

ASRDeepspeech x Sakura-ML (English/Japanese)

Modules • Code structure • Installing the application • Makefile commands • Environments • Dataset• Running the application• Notes•

This repository offers a clean code version of the original repository from SeanNaren with classes and modular components (eg trainers, models, loggers...).

I have added a configuration file to manage the parameters set in the model. You will also find a pretrained model in japanese performing a CER = 34 on JSUT test set .

Modules

At a granular level, ASRDeepSpeech is a library that consists of the following components:

Component	Description
asr_deepspeech	Speech Recognition package
asr_deepspeech.data	Data related module
asr_deepspeech.data.dataset	Build the dataset
asr_deepspeech.data.loaders	Load the dataset
asr_deepspeech.data.parsers	Parse the dataset
asr_deepspeech.data.samplers	Sample the dataset
asr_deepspeech.decoders	Decode the generated text
asr_deepspeech.loggers	Loggers
asr_deepspeech.modules	Components of the network
asr_deepspeech.parsers	Arguments parser
asr_deepspeech.tests	Test units
asr_deepspeech.trainers	Trainers

Code structure

from setuptools import setup
from asr_deepspeech import __version__

setup(
    name="asr_deepspeech",
    version=__version__,
    short_description="ASRDeepspeech (English / Japanese)",
    long_description="".join(open("README.md", "r").readlines()),
    long_description_content_type="text/markdown",
    url="https://github.com/zakuro-ai/asr",
    license="MIT Licence",
    author="CADIC Jean-Maximilien",
    python_requires=">=3.8",
    packages=[
        "asr_deepspeech",
        "asr_deepspeech.audio",
        "asr_deepspeech.data",
        "asr_deepspeech.data.dataset",
        "asr_deepspeech.data.loaders",
        "asr_deepspeech.data.manifests",
        "asr_deepspeech.data.parsers",
        "asr_deepspeech.data.samplers",
        "asr_deepspeech.decoders",
        "asr_deepspeech.etl",
        "asr_deepspeech.loggers",
        "asr_deepspeech.models",
        "asr_deepspeech.modules",
        "asr_deepspeech.parsers",
        "asr_deepspeech.tests",
        "asr_deepspeech.trainers",
    ],
    include_package_data=True,
    package_data={"": ["*.yml"]},
    install_requires=[r.rsplit()[0] for r in open("requirements.txt")],
    author_email="[e-mail]",
    description="ASRDeepspeech (English / Japanese)",
    platforms="linux_debian_10_x86_64",
    classifiers=[
        "Programming Language :: Python :: 3.8",
        "License :: OSI Approved :: MIT License",
    ],
)

Installing the application

To clone and run this application, you'll need the following installed on your computer:

• Git
• Docker Desktop
  • Install Docker Desktop on Mac
  • Install Docker Desktop on Windows
  • Install Docker Desktop on Linux
• Python

Install bpd:

# Clone this repository and install the code
git clone https://github.com/zakuro-ai/asr

# Go into the repository
cd asr

Makefile commands

Exhaustive list of make commands:

pull                # Download the docker image
sandbox             # Launch the sandox image 
install_wheels      # Install the wheel
tests               # Test the code

Environments

We are providing a support for local or docker setup. However we recommend to use docker to avoid any difficulty to run the code. If you decide to run the code locally you will need python3.6 with cuda>=10.1. Several libraries are needed to be installed for training to work. I will assume that everything is being installed in an Anaconda installation on Ubuntu, with Pytorch 1.0. Install PyTorch if you haven't already.

Docker

Note

Running this application by using Docker is recommended.

To build and run the docker image

make pull
make sandbox

PythonEnv

Warning

Running this application by using PythonEnv is possible but not recommended.

make install_wheels

Test

make tests

You should be able to get an output like

=1= TEST PASSED : asr_deepspeech
=1= TEST PASSED : asr_deepspeech.data
=1= TEST PASSED : asr_deepspeech.data.dataset
=1= TEST PASSED : asr_deepspeech.data.loaders
=1= TEST PASSED : asr_deepspeech.data.parsers
=1= TEST PASSED : asr_deepspeech.data.samplers
=1= TEST PASSED : asr_deepspeech.decoders
=1= TEST PASSED : asr_deepspeech.loggers
=1= TEST PASSED : asr_deepspeech.modules
=1= TEST PASSED : asr_deepspeech.parsers
=1= TEST PASSED : asr_deepspeech.test
=1= TEST PASSED : asr_deepspeech.trainers

Datasets

By default we process the JSUT dataset. See the config section to know how to process a custom dataset.

from gnutools.remote import gdrive
from asr_deepspech import  cfg

# This will download the JSUT dataset in your /tmp
gdrive(cfg.gdrive_uri)

ETL

python -m asr_deepspeech.etl

Running the application

Training a Model

To train on a single gpu

sakura -m asr_deepspeech.trainers

Pretrained model

python -m asr_deepspeech

Notes

Clean verbose during training

================ VARS ===================
manifest: clean
distributed: True
train_manifest: __data__/manifests/train_clean.json
val_manifest: __data__/manifests/val_clean.json
model_path: /data/ASRModels/deepspeech_jp_500_clean.pth
continue_from: None
output_file: /data/ASRModels/deepspeech_jp_500_clean.txt
main_proc: True
rank: 0
gpu_rank: 0
world_size: 2
==========================================

Progress bar

...
clean - 0:00:46 >> 2/1000 (1) | Loss 95.1626 | Lr 0.30e-3 | WER/CER 98.06/95.16 - (98.06/[95.16]): 100%|██████████████████████| 18/18 [00:46<00:00,  2.59s/it]
clean - 0:00:47 >> 3/1000 (1) | Loss 96.3579 | Lr 0.29e-3 | WER/CER 97.55/97.55 - (98.06/[95.16]): 100%|██████████████████████| 18/18 [00:47<00:00,  2.61s/it]
clean - 0:00:47 >> 4/1000 (1) | Loss 97.5705 | Lr 0.29e-3 | WER/CER 100.00/100.00 - (98.06/[95.16]): 100%|████████████████████| 18/18 [00:47<00:00,  2.66s/it]
clean - 0:00:48 >> 5/1000 (1) | Loss 97.8628 | Lr 0.29e-3 | WER/CER 98.74/98.74 - (98.06/[95.16]): 100%|██████████████████████| 18/18 [00:50<00:00,  2.78s/it]
clean - 0:00:50 >> 6/1000 (5) | Loss 97.0118 | Lr 0.29e-3 | WER/CER 96.26/93.61 - (96.26/[93.61]): 100%|██████████████████████| 18/18 [00:49<00:00,  2.76s/it]
clean - 0:00:50 >> 7/1000 (5) | Loss 97.2341 | Lr 0.28e-3 | WER/CER 98.35/98.35 - (96.26/[93.61]):  17%|███▊                   | 3/18 [00:10<00:55,  3.72s/it]
...

Separated text file to check wer/cer with histogram on CER values (best/last/worst result)

================= 100.00/34.49 =================
----- BEST -----
Ref:良ある人ならそんな風にに話しかけないだろう
Hyp:用ある人ならそんな風にに話しかけないだろう
WER:100.0  - CER:4.761904761904762
----- LAST -----
Ref:すみませんがオースチンさんは5日にはです
Hyp:すみませんがースンさんは一つかにはです
WER:100.0  - CER:25.0
----- WORST -----
Ref:小切には内がみられる
Hyp:コには内先金地つ作みが見られる
WER:100.0  - CER:90.0
CER histogram
|###############################################################################
|███████████                                                           6  0-10  
|███████████████████████████                                          15  10-20 
|███████████████████████████████████████████████████████████████████  36  20-30 
|█████████████████████████████████████████████████████████████████    35  30-40 
|██████████████████████████████████████████████████                   27  40-50 
|█████████████████████████████                                        16  50-60 
|█████████                                                             5  60-70 
|███████████                                                           6  70-80 
|                                                                      0  80-90 
|█                                                                     1  90-100
=============================================

Acknowledgements

Thanks to Egor and Ryan for their contributions!

This is a fork from https://github.com/SeanNaren/deepspeech.pytorch. The code has been improved for the readability only.

For any question please contact me at j.cadic[at]protonmail.ch