appautomaton/openmoss-audio-tokenizer-mlx

OpenMOSS Audio Tokenizer — MLX 8-bit

This repository contains an MLX-native int8 conversion of the OpenMOSS audio tokenizer for Apple Silicon.

It is a supporting model that encodes and decodes audio tokens for the OpenMOSS TTS family. It is not a standalone speech generation model.

Variants

Path	Precision
mlx-int8/	int8 quantized weights

Model Details

• Developed by: AppAutomaton
• Shared by: AppAutomaton on Hugging Face
• Upstream model: OpenMOSS-Team/MOSS-Audio-Tokenizer
• Task: audio tokenization and codec decoding
• Runtime: MLX on Apple Silicon

How to Get Started

Load it directly with mlx-speech:

from mlx_speech.models.moss_audio_tokenizer import MossAudioTokenizerModel

model = MossAudioTokenizerModel.from_path("mlx-int8")

The tokenizer is loaded automatically when you run OpenMOSS generation scripts. You usually do not need to instantiate it directly.

python scripts/generate_moss_local.py \
  --text "Hello from mlx-speech." \
  --output outputs/out.wav

Notes

• This repo contains the quantized MLX runtime artifact only.
• The conversion remaps the original OpenMOSS audio tokenizer weights explicitly for MLX inference.
• The artifact is shared by the OpenMOSS local TTS, TTSD, and SoundEffect runtime paths in this repo.

Links

• Source code: mlx-speech
• More examples: AppAutomaton

License

Apache 2.0 — following the upstream license published with OpenMOSS-Team/MOSS-Audio-Tokenizer.