microsoft/UniSpeech

Unifies self-supervised and supervised learning across multiple speech tasks (ASR, speaker recognition, speech enhancement) through variants like WavLM and UniSpeech-SAT that incorporate speaker-aware pre-training and intermediate layer supervision. Models scale from 960 hours (LibriSpeech) to 94k hours across Libri-Light, GigaSpeech, and VoxPopuli datasets, with multilingual support for English, French, Spanish, and Italian. Fully integrated with HuggingFace for straightforward model loading and fine-tuning on downstream tasks.

479 stars. No commits in the last 6 months.