metavoiceio/metavoice-src

Combines a causal GPT-based LLM for EnCodec token prediction with multi-band diffusion and DeepFilterNet post-processing to generate waveforms, supporting zero-shot voice cloning from 30s reference audio and cross-lingual finetuning with minimal data. Optimized for sub-real-time inference on modern GPUs via torch.compile and KV-caching, deployable via Docker, Hugging Face, or standalone Python APIs with optional int4/int8 quantization modes. Supports arbitrary-length synthesis and finetuning on custom speaker datasets via CSV-formatted audio/caption pairs, with optional W&B integration for experiment tracking.

4,201 stars. No commits in the last 6 months.