asr_project

Automatic Speech Recognition (ASR) Project

This repository contains the implementation of the DeepSpeech2 model for the ASR task. To reproduce the final model's result train your model with the src/configs/deepspeech2_augs.json configuration for 75 epochs.

Results

WER = $0.276992$ and CER = $0.125911$ were achieved on test-other split by the final model (using beam-search + language model).

Installation

• Clone this repository

git clone https://github.com/hzchet/asr_project.git
cd asr_project

• In order to install all the required packages, run the following command in your terminal:

  pip install -r requirements.txt

• Install language models (used for beam-search rescoring) by running the following commands

wget -P ./saved/lms/ https://www.openslr.org/resources/11/3-gram.arpa.gz
wget -P ./saved/lms/ https://www.openslr.org/resources/11/3-gram.arpa.gz
gzip -d ./saved/lms/3-gram.arpa.gz
gzip -d ./saved/lms/4-gram.arpa.gz

• Install the weights of the pre-trained model by running

python3 install_weights.py

• Copy the config into the same directory

cp src/configs/deepspeech2_augs.json saved/models/final/config.json

Tests

In order to run unit tests run the following command

python3 unit_tests.py

Metrics

In order to reproduce metrics on test-clean/test-other datasets, run the following command

python3 test.py -r saved/models/final/weights.pth