WeNet is a production-ready, end-to-end speech recognition toolkit designed for both streaming and non-streaming applications. It offers a full-stack solution for ASR development, targeting researchers and engineers who need accurate, lightweight, and well-documented tools for building and deploying speech recognition systems.

How It Works

WeNet integrates both Transformer and Conformer models, leveraging a hybrid approach that combines the strengths of different architectures for state-of-the-art accuracy. It supports WFST-based decoding for seamless Language Model integration and offers efficient runtime solutions for deployment.

Quick Start & Requirements

• Install (runtime only): pip install git+https://github.com/wenet-e2e/wenet.git
• Install (training/deployment): Clone repo, create Conda env (conda create -n wenet python=3.10), install sox (conda install conda-forge::sox), PyTorch (pip install torch==2.2.2+cu121 torchaudio==2.2.2+cu121 -f https://download.pytorch.org/whl/torch_stable.html), and other dependencies (pip install -r requirements.txt).
• Prerequisites: CUDA 12.1 recommended, Python 3.10, sox and libsox-dev (Ubuntu/CentOS). Ascend NPU support requires CANN toolkit.
• Runtime build: Requires cmake 3.14+.
• Docs: Roadmap, Docs

Highlighted Details

• Achieves SOTA results on public speech datasets.
• Production-first design with full-stack solutions.
• Supports both streaming and non-streaming ASR.
• Well-documented with Python and command-line usage examples.

Maintenance & Community

• Active development with contributions from multiple authors.
• Discussion primarily via GitHub Issues. WeChat group available for Chinese users.

Licensing & Compatibility

• Licensed under Apache 2.0.
• Compatible with commercial use and closed-source linking.

Limitations & Caveats

The runtime build for x86 or LM integration requires manual compilation steps. Specific hardware acceleration (e.g., Ascend NPU) necessitates separate installation of vendor-specific toolkits and kernel drivers.