balisujohn/tortoise.cpp

Enables multi-speaker voice synthesis via speaker embedding files, with inference optimized across CPU, CUDA, and Metal backends. The architecture chains three GGML models—autoregressive, vocoder, and diffusion—with deterministic seed control for reproducible audio generation. Custom voices are extracted from the original tortoise-tts pipeline and serialized as binary conditioning files for efficient inference.

192 stars. No commits in the last 6 months.