AlexandaJerry/vits-mandarin-biaobei

Implements conditional VAE with adversarial learning and normalizing flows for end-to-end Mandarin speech synthesis, enabling parallel decoding with stochastic duration prediction to capture natural variations in pitch and rhythm. Trained on the Biaobei dataset with phoneme-level text processing via g2p conversion and monotonic alignment search. Supports single-speaker and multi-speaker configurations through separate training pipelines optimized for 22050 Hz audio.

121 stars. No commits in the last 6 months.