Icefall
The recipe book for next-gen Kaldi: reproducible Conformer and Zipformer ASR training across 20+ datasets.
About
Icefall is the recipe collection for next-gen Kaldi: reproducible, open training recipes for Conformer and Zipformer ASR across more than 20 datasets, from LibriSpeech and GigaSpeech to AISHELL and Common Voice. It covers CTC, LF-MMI, and transducer training, streaming and non-streaming, plus Whisper fine-tuning and VITS text-to-speech recipes.
It is a research and production-training tool rather than a model you download: you train your own ASR with k2 and Lhotse, then export to ONNX for deployment through sherpa-onnx. Teams like Reazon Research have used it to build streaming Japanese models. If you want to train custom ASR on your own data with full control, Icefall is the most complete open starting point.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Recipe-level results published per recipe (e.g. 0.42% WER on the yesno toy recipe; LibriSpeech recipe tables, provider-reported)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: Apache-2.0 repo from the k2-fsa org with 20+ dataset recipes documented in the project README; Zipformer/Conformer transducer, CTC, and MMI coverage confirmed across forks and integrator docs; Reazon Research's streaming Japanese build as real-world adoption. Strengths: the most complete open ASR training recipe set, direct pipeline into sherpa-onnx deployment, active use by research teams. Docked for: (1) it is a training toolkit, not a downloadable model, so entry-level value is lower; (2) results are recipe-level and provider-reported with no single headline metric. License: Apache-2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Training recipe collection, not a downloadable model: Conformer/Zipformer CTC, MMI, transducer across 20+ datasets
- · Exports to ONNX for sherpa-onnx deployment; used by Reazon Research for streaming Japanese models