istupakov/onnx-asr

Supports multiple modern ASR architectures (NeMo Conformer/Parakeet/Canary, GigaAM, Kaldi Icefall Zipformer, Whisper) with built-in preprocessing and greedy decoding, eliminating the need for external dependencies like PyTorch or FFmpeg. Runs on heterogeneous hardware via ONNX Runtime backends (CUDA, TensorRT, CoreML, DirectML, ROCm), from edge devices to GPUs, with VAD-based long-form recognition and token-level timestamps. Integrates with Hugging Face Hub for model distribution and accepts WAV files or NumPy arrays with automatic resampling.

281 stars and 20,143 monthly downloads. Available on PyPI.