silentz/Deep-Speech-2-End-To-End-Speech-Recognition-In-English-And-Mandarin

DeepSpeech2: https://arxiv.org/abs/1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

LibriSpeech CER test-clean: 0.260732

LibriSpeech WER test-clean: 0.618628

How to reproduce model

1. Setup virtualenv

Execute following commands from project root:

virtualenv --python=python3.8 venv
source venv/bin/activate
pip install pybind11
pip install -r requirements.txt

git clone --recursive https://github.com/parlance/ctcdecode.git
cd ctcdecode && pip install . && cd ..

2. Download artifacts (model weights and LM)

Execute following commands from project root:

./download_artifacts.sh

3. Launch inference on librispeech test-clean

From project root:

./test.sh

4. Learn model from scratch

From project root:

./train.sh

Training logs

Stage 1 (LJSpeech)

Stage 2 (Librispeech train-clean-360)