---
license: apache-2.0
library_name: mlx
tags:
  - text-to-speech
  - voice-cloning
  - mlx
  - omnivoice
pipeline_tag: text-to-speech
---

# OmniVoice-MLX-int8

8-bit quantized [MLX](https://github.com/ml-explore/mlx) conversion of
[k2-fsa/OmniVoice](https://github.com/k2-fsa/OmniVoice) — a single-stage,
non-autoregressive discrete-diffusion text-to-speech model with a Qwen3 backbone
and broad multilingual coverage (600+ languages). For Apple Silicon, via the
`speech-swift` runtime.

## Contents

| File | Precision | Size |
|---|---|---|
| `model.safetensors` | int8 backbone (group 64, 8-bit) | ~689 MB |
| `audio_tokenizer/model.safetensors` | fp16 codec (Higgs-audio v2) | ~403 MB |
| `tokenizer.json`, `config.json`, … | — | — |

The diffusion backbone is quantized (it dominates latency — 32 forward passes per
clip); the codec stays fp16 since it runs once and is already exact there.

## Why int8

The backbone is memory-bandwidth bound at these sequence lengths, so 8-bit weights
cut latency with no audible loss. On an M5 Pro: **RTF ≈ 0.18** (≈ 5.5× faster than
real-time), versus ≈ 0.25 at fp32, and the ASR roundtrip stays at **0% WER**.

Precision tiers:

| Variant | Backbone | Total | Use |
|---|---|---|---|
| `OmniVoice-MLX-fp16` | fp16 | ~1.5 GB | balanced |
| **`OmniVoice-MLX-int8`** | **int8** | **~1.0 GB** | **fastest / smallest** |

## Usage

Loaded by the `OmniVoiceTTS` module in `speech-swift`. The `quantization` block in
`config.json` (`group_size: 64`, `bits: 8`) tells the loader to swap the
Linear/Embedding layers to their quantized form before loading the weights.

## License

Apache-2.0, following the upstream model.