Deep Speech with CTC Loss

Introduction

Deep Speech model is one of the ASR that got the SOTA in Speech Recognition domain. In this respository, I use Deep Speech with Vivos Dataset and Vin BigData VLSP 2020 Dataset.

How to use this respository

1. Clone this project to current directory. Using those commands:

!git init
!git remote add origin https://github.com/tuanio/deepspeech-ctc
!git pull origin main

2. Install requirement packages

!pip install -r requirements.txt

Then install ctcdecode from this respository: https://github.com/parlance/ctcdecode

3. Edit configs.yaml file for appropriation.
4. Train model using python main.py -cp conf -cn configs

Run the Web Demo version

• streamlit run web.py

Train results

Train loss of Deep Speech on 978 epochs

Validation loss of Deep Speech

Validation word error rate (mean wer) of Deep Speech

Note

• sox is audio backend for linux, PySoundFile is audio backend for windows

Environment variable

• HYDRA_FULL_ERROR=1