yl4579/PL-BERT

Implements a phoneme-level ALBERT encoder trained with dual pretext tasks—masked phoneme prediction and grapheme-to-phoneme correspondence—enabling efficient prosodic feature extraction for downstream TTS models. The architecture operates on phoneme sequences rather than word-level inputs, eliminating inefficiencies when integrated into phoneme-based synthesis pipelines like StyleTTS. Pre-trained checkpoints target English Wikipedia data and plug directly into existing TTS frameworks via a linear projection layer that bridges hidden dimension mismatches.

268 stars. No commits in the last 6 months.