Rybib/chatterbox-nano-fp16-mlx

Chatterbox Nano FP16 MLX

A native FP16 MLX conversion of ResembleAI/chatterbox-nano, prepared by Rybib for Apple Silicon and on-device experimentation.

Chatterbox Nano and its architecture were created by Resemble AI. This is an independent conversion, not an official Resemble AI release.

Package

• Native MLX tensor names and layouts
• FP16 floating-point weights; no 4-bit quantization
• Nano's distilled mean-flow speech decoder
• English tokenizer and default voice conditioning
• Approximately 697 MB for speech generation
• Bundled S3TokenizerV2 companion model for offline voice cloning
• Approximately 1.16 GB total with voice cloning

The converter strictly verifies that every required Nano decoder and vocoder parameter is mapped. It also merges PyTorch weight-normalization parameters into inference weights and converts convolution/GPT-2 layouts for MLX.

Quality

Listening tests on Apple Silicon found the output perceptually equivalent to the official FP32 Nano model. This is not a formal benchmark; applications should perform their own voice-cloning and long-form tests.

Usage

This package requires a Chatterbox Nano-capable MLX Audio runtime. At the time of conversion, upstream MLX Audio did not correctly map all Nano mean-flow decoder tensors without the accompanying runtime/converter fixes.

Voice cloning additionally requires the S3TokenizerV2 speech tokenizer. It is included under S3TokenizerV2/ so applications can clone voices completely offline after one installation.

Provenance and license

• Base model: ResembleAI/chatterbox-nano
• Upstream code: resemble-ai/chatterbox
• MLX ecosystem: Blaizzy/mlx-audio
• Speech tokenizer: mlx-community/S3TokenizerV2
• Conversion precision: IEEE FP16
• Conversion owner: Rybib

The upstream Chatterbox model is distributed under the MIT License. Its copyright notice is preserved in LICENSE.