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.
