jaywalnut310/vits

Combines normalizing flows with adversarial training to enable parallel, single-stage synthesis that matches two-stage TTS quality while modeling natural speech variation through a stochastic duration predictor. Implements monotonic alignment search (Cython-optimized) for unsupervised duration learning and supports both single-speaker (LJ Speech) and multi-speaker (VCTK) training pipelines with PyTorch, requiring phoneme preprocessing via g2p.

7,837 stars. No commits in the last 6 months.