anyvoiceai/Barkify

Implements a two-stage GPT-style training pipeline for Bark's text-to-speech model: semantic token prediction (text→semantic) followed by acoustic modeling (semantic→audio). Supports multi-speaker scenarios with speaker and audio prompting capabilities, trained on datasets like LJSpeech and LibriTTS using standard PyTorch workflows configured via YAML. Provides inference through Jupyter notebooks and demonstrates compatibility with variable-length audio and cross-lingual experiments.

130 stars. No commits in the last 6 months.