k2-fsa/icefall

Health v2 · maintenance only

Adoption not part of the score

1482 stars · 426 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

Icefall is a collection of speech recognition (ASR) and TTS training recipes built on the k2 and lhotse libraries, implemented in Python with PyTorch. It provides recipes for many public datasets and models such as Zipformer, Conformer, Transducer, CTC, and Whisper, with export paths to deployment frameworks like sherpa-onnx.

Use cases

• train an automatic speech recognition model on librispeech
• fine-tune whisper on a custom speech dataset
• train a zipformer transducer ASR model
• build a TTS model with VITS on LJSpeech
• export a trained ASR model to onnx for deployment
• run forced alignment on speech transcripts
• reproduce ASR benchmarks like gigaspeech or wenetspeech

When to choose

• you want to train state-of-the-art ASR models like Zipformer or Conformer transducers
• you need reproducible recipes for standard speech datasets
• you plan to deploy with the sherpa family of runtimes
• you want FST-based decoding, alignment, or language model rescoring with k2

When to avoid

• you only need a ready-made inference engine rather than training recipes
• you want a simple pip-installable ASR API without GPU training setup
• you work outside speech/audio domains
• you need Windows-native training support without Docker

Facets

library · maturity active

speech-recognition machine-learning deep-learning llm-training speech-processing machine-learning deep-learning python asr k2 lhotse speech-recipes tts zipformer transducer model-training pytorch linux gpu docker

Member repositories

For agents

markdown · JSON · MCP: product_card(name="k2-fsa/icefall")