ASR Project

Project Name

ASR Project is a production-oriented workspace for building custom Automatic Speech Recognition (ASR) models and related audio ML tooling.

Short Description

This repository provides a practical foundation for end-to-end speech ML workflows:

• dataset preparation and versioning;
• feature engineering and model training;
• evaluation with ASR metrics (WER/CER and related checks);
• reproducible experiments and inference-ready packaging.

Project Structure

• src/asr/ - application source code (data, features, models, training, inference, utils).
• data/ - local datasets (raw, interim, processed, external).
• notebooks/ - exploratory and experiment notebooks.
• configs/ - experiment and pipeline configuration files.
• scripts/ - CLI and automation scripts.
• models/, checkpoints/, artifacts/, logs/ - training outputs and runtime artifacts.
• tests/ - unit and integration tests.
• docs/ - project and code documentation in multiple languages.

Documentation

Documentation is organized by language and purpose.

Language support:

• docs/en/ - English documentation.
• docs/ru/ - Russian documentation.

Detailed dataset installation and downloader docs:

• English: docs/en/code-documentation/scripts_data_download.md
• Russian: docs/ru/code-documentation/scripts_data_download.md

Installation

Requirements:

• Python 3.12+
• Poetry

Install project dependencies:

poetry install --with dev

Install ML stack when needed:

poetry install --with dev,ml

Usage

Run quality checks:

poetry run ruff check .
poetry run mypy src tests
poetry run pytest

Quick dataset download example:

bash scripts/data_download.sh --datasets all

Install SOVA from local pre-downloaded archive:

bash scripts/data_download.sh --sova-archive /path/to/RuDevices.tar

Source for manual SOVA archive download:

• https://disk.yandex.ru/d/jz3k7pnzTpnTgw

Resume interrupted or verify existing dataset downloads:

bash scripts/data_download.sh --datasets golos

Force full re-download of built-in downloadable datasets:

bash scripts/data_download.sh reinstall-all

Retry policy overview:

• request-level retries: 5
• dataset-level retries: 3
• global pipeline retries: 5

Start notebooks:

poetry run jupyter lab

Pre-commit

Pre-commit is configured for ruff, mypy, and pytest.

Setup:

poetry run pre-commit install
poetry run pre-commit run --all-files

Testing

Run all tests:

poetry run pytest

Run integration tests only:

poetry run pytest -m integration

Contributing

Contribution guide: CONTRIBUTING.md.

Security

Security policy and responsible disclosure contacts: SECURITY.md.

License

This project is licensed under Apache License 2.0. See LICENSE.