UnknownResearcher/chatterbox-turbo-coreml

Chatterbox Turbo, Core ML

Resemble AI's Chatterbox Turbo converted to Core ML for Apple Silicon, laid out like FluidInference/chatterbox-nano-coreml so the same pipeline runs it.

Turbo and Nano share the S3Gen decoder, voice encoder and tokenizer byte for byte, so FlowMean-N500, HiFT-T1000 and tokenizer/ are copied unchanged from the Nano repo. Only the T3 token model (GPT2-medium: 24 layers, 1024 wide, 16 heads) and tables/ are new.

File	What it is
T3Turbo-Prefill-T512-M1536-fp16.mlmodelc	T3 prefill over voice conditioning, text and BOS
T3Turbo-Decode-M1536-fp16-stateful.mlmodelc	T3 decode step, KV cache in MLState (macOS 15+)
FlowMean-N500-fp16.mlmodelc, HiFT-T1000-fp16.mlmodelc	S3Gen meanflow decoder and HiFT vocoder, from the Nano repo
tables/tables.safetensors	Turbo's text and speech embedding tables (fp16)
tables/voice-default.safetensors	Turbo's built-in voice from conds.pt

Converted with FluidInference's mobius Nano scripts adapted to GPT2-medium. Against PyTorch, the fp16 Core ML logits differ by at most 0.014 (prefill) and 0.016 (stateful decode).

Run it with the ChatterboxNanoVariant.turbo patch to FluidAudio at jordan-schnur/FluidAudio.

Model weights © Resemble AI, MIT licence. Generated audio has no Perth watermark, since the Core ML pipeline doesn't apply one.