aufklarer/Chatterbox-Multilingual-MLX-fp16

Multilingual Chatterbox (Resemble AI, MIT) converted to MLX in genuine fp16 for Apple-silicon inference. Unlike mlx-community/chatterbox-fp16 (stored as F32, 2.6 GB), this bundle stores all float tensors as fp16 (1.3 GB) with an identical key layout.

Zero-shot voice cloning from a short reference clip across 23 languages (incl. Arabic and Hindi). Component prefixes: ve.* (voice encoder), t3.* (text→speech-token T3), s3gen.* (token→waveform S3Gen).

Note: requires the S3Tokenizer weights from mlx-community/S3TokenizerV2, downloaded automatically at runtime.

Use with mlx-audio

pip install -U mlx-audio
mlx_audio.tts.generate --model aufklarer/Chatterbox-Multilingual-MLX-fp16 --text "[ar] مرحبا" --ref_audio reference.wav

Converted with models/chatterbox/export/convert.py (speech-models).