Kaldi recipes that feed frozen self-supervised encoder features (HuBERT, WavLM, MMS, mHuBERT, AV-HuBERT, etc.) to GMM-HMM, TDNN-F chain and flat-start e2e chain models. Multilingual ASR recipes, 3.88% WER on Mini LibriSpeech, and GRID lipreading at 5.85% WER on unseen speakers.
multilingual speech-recognition speech-to-text stt asr visual-speech-recognition kaldi-asr lf-mmi self-supervised-learning low-resource-languages lipreading ssl-features
-
Updated
Oct 21, 2026 - Shell