icefall

A speech recognition toolkit for training and adapting models locally, with streaming recipes, CPU and CUDA Docker support, and ONNX export.

Screenshot of icefall website

icefall is a speech model training toolkit for developers and researchers building recognition systems on their own hardware. Its Python recipes use k2 and lhotse to train, fine-tune, and evaluate models across speech datasets.

  • Speech recognition: Train Zipformer and Conformer models for streaming or complete recordings, with recipes for LibriSpeech, GigaSpeech, CommonVoice, Aishell, and meeting speech datasets.
  • Model adaptation: Fine-tune pretrained Zipformer models, including with adapters, or fine-tune Whisper on AiShell-1.
  • Alignment and language models: Match speech to transcripts with forced alignment, train RNN language models, and use language models to rescore recognition results.
  • Deployment exports: Export trained models to TorchScript, ONNX, or ncnn for use with the related sherpa, sherpa-ncnn, and sherpa-onnx runtimes.
  • Speech generation: Train VITS text-to-speech models with recipes for LJSpeech, VCTK, and LibriTTS.

icefall is open source under Apache 2.0. It provides CPU and CUDA Docker images; the small yesno recognition recipe can train on a CPU. Pretrained models are available through Hugging Face, and hosted Hugging Face Spaces provide browser demos separate from the toolkit running on your hardware.

Similar to icefall