OpenASR-py

OpenASR-py is a minimal, PyTorch based open source toolkit for end-to-end automatic speech recognition (ASR) related tasks, which borrows many elements from OpenNMT-py and at the same time provides simpler, task-specific reimplementations of several others.

Source: https://www.clsp.jhu.edu/workshops/18-workshop/multilingual-end-end-asr-incomplete-data/

Due to the highly modular and transparent codebase, it can be used as a starting point for research projects in ASR as well as other less explored topics such as domain adaptation, adversarial training, active learning etc.

Key features

• Blazingly fast, just like OpenNMT-py (details here)
• Highly modular and easily extensible codebase
• Provision of basic routine(s) for ASR
• Audio-specific feature extraction and data preprocessing
• Simple and transparent data loading pipeline
• Implementations of a variety of encoders, decoders and attention mechanisms
• Support for word-, character- and wordpiece-level output granularity
• Beam-search decoding and error rate computation during evaluation
• Logging support using Tensorboard
• Model checkpointing and resumable training

Contents

1. Installation
2. Overview
3. Pipelines
4. Acknowledgements
5. Contributing

Installation

We recommend using conda for setting up the environment. After it has been successfully installed, follow the steps below:

# Create environment
conda create -n oasr
conda activate oasr
# Install pytorch 1.1 and its dependencies
# NOTE: This command gives an intermittent 'HTTP 000 Connection Error'. 
# Retrying it, several times at worst, solves the issue.
conda install pytorch cudatoolkit=10.0 -c pytorch
# Clone codebase and install its dependencies
git clone https://github.com/csalt-research/OpenASR-py.git
cd OpenASR-py/
pip install -r requirements.txt

Overview

TODO

Pipelines

We provide functional code for the following tasks. You can find more details in the corresponding README files.

1. Automatic Speech Recognition (ASR): obtain the transcription for a given utterance
2. Domain Adversarial Training (DAT): TODO
3. Active Learning (AL): TODO
4. Active Adversarial Domain Adaptation (AADA): TODO

Acknowledgements

OpenASR-py was originally developed by Yash Shah (ys1998) using the OpenNMT-py framework as a starting point; it was initiated with the objective of making certain relatively complicated and opaque aspects of OpenNMT-py more ASR-specific and research friendly during his undergraduate thesis project at IIT Bombay under the supervision of Prof. Preethi Jyothi.

Contributing

Feel free to report any bug, request a feature or ask a general question in the Issues tab. We also love contributions, for which you can consult the same section for appropriately tagged posts.