santi-pdp/pase

Self-supervised speech encoder trained with a worker/minion multi-task framework, where auxiliary "worker" networks guide encoder learning across diverse speech tasks (ASR, speaker recognition, emotion detection, voice conversion). Encodes raw waveforms into fixed-dimensional frame representations (256-dim, 625 frames for 100k samples) that transfer directly to downstream tasks via fine-tuning. Implements data augmentation with codec2-based distortions, configurable learning rate scheduling (polynomial/step decay), and validation on held-out speakers to ensure generalization without labeled data.

447 stars. No commits in the last 6 months.