coqui-ai/TTS

Supports multiple model architectures spanning spectrogram-based (Tacotron2, Glow-TTS, FastSpeech2) and end-to-end approaches (VITS, XTTS), with built-in speaker encoder for multi-speaker synthesis and voice cloning. Enables sub-200ms streaming inference, fine-tuning on custom datasets, and integrates ~1100 Fairseq models alongside modular vocoder support (MelGAN, ParallelWaveGAN, WaveGrad). Training infrastructure includes dataset curation tools, Tensorboard logging, and a lightweight Trainer API optimized for efficient multi-GPU training.

44,801 stars and 214,937 monthly downloads. Used by 2 other packages. No commits in the last 6 months. Available on PyPI.