gokhaneraslan/chatterbox-finetuning

The toolkit supports dual architectures—Llama-based Standard mode with grapheme tokenization for language fundamentals, and GPT-2-based Turbo mode that automatically merges a 50K+ token BPE vocabulary with multi-language grapheme sets for accelerated fine-tuning. Preprocessing is mandatory and offline, extracting speaker embeddings and acoustic tokens upfront to maximize training throughput; the pipeline handles automatic resampling to 16kHz training input and generates 24kHz output via vocoder, with integrated VAD support for inference trimming.