jaywalnut310/glow-tts

Combines normalizing flows with dynamic programming-based monotonic alignment search to enable parallel mel-spectrogram generation without requiring external aligners, eliminating the dependency on autoregressive teacher models. Integrates with HiFi-GAN vocoder for improved audio quality and supports multi-speaker synthesis through conditional generation. Achieves order-of-magnitude speedup over Tacotron 2 while maintaining comparable speech quality with controllable and diverse output.

704 stars. No commits in the last 6 months.