wenet-e2e/wenet

WeNet is an end-to-end speech recognition toolkit built with PyTorch that prioritizes production-ready deployment alongside research capabilities.

The toolkit addresses the need for speech recognition systems that work reliably in production environments while maintaining state-of-the-art accuracy. WeNet combines multiple architectural approaches including Conformer and Transformer models, supporting both training from scratch and inference with pretrained models. It handles multiple languages through different model configurations and provides both command-line and Python API interfaces for integration into applications.

WeNet suits teams building speech recognition systems who need production deployment as a first-class concern rather than an afterthought. The toolkit works well for projects requiring Chinese speech recognition through models like Paraformer and WenetSpeech, or English recognition via Whisper integration. Installation is straightforward through pip for inference-only use cases, though GPU usage requires explicit configuration and careful attention to CUDA version compatibility with your driver. For training and deployment workflows, the project supports both standard x86 runtime and Ascend NPU hardware through separate installation paths, with detailed version compatibility tables provided for NPU users.

The project maintains active issue-based discussion channels and provides comprehensive documentation including papers, pretrained models on HuggingFace, and a dedicated guru chatbot for questions. Development activity shows consistent attention to production concerns, with explicit GPU hardware notes addressing common pitfalls like silent CPU fallback and dtype considerations for specific accelerators. The codebase includes specialized support for multiple hardware platforms beyond standard GPUs, indicating ongoing work to broaden deployment options. Runtime components are maintained as a separate concern with platform-specific build documentation, reflecting a deliberate separation between training infrastructure and production serving.