---
license: other
license_name: boson-ai-community
license_link: LICENSE
base_model: bosonai/higgs-audio-v3-tts-4b
base_model_relation: quantized
pipeline_tag: text-to-speech
tags:
- higgs-audio
- text-to-speech
- onnx
- onnxruntime
- int4
- quantized
- vocoder
---

# Higgs Audio v3 TTS 4-bit ONNX Vocoder

This repository contains an ONNX export of the Higgs Audio v3 bundled codec
decoder/vocoder:

```text
audio_codes [B, 8, T] int64 -> audio_values [B, 1, samples] float32
```

It is a browser-deployment component for Higgs Audio TTS, not the complete
text-to-speech pipeline. The autoregressive Qwen3 text-to-codebook generator
is still a separate custom SGLang graph and is not included here.

## Files

- `higgs_audio_v3_vocoder_decode.onnx` + `.onnx.data`: fp32 ONNX vocoder.
- `higgs_audio_v3_vocoder_decode_matmul4.onnx` + `.onnx.data`: ONNX Runtime
  weight-only 4-bit quantization for supported `MatMul` and `Gather` nodes.
- `vocoder_onnx_export_report.json`: PyTorch vs ONNX Runtime validation.
- `vocoder_onnx_matmul4_report.json`: 4-bit ONNX Runtime validation.

## Validation

The fp32 vocoder export passed ONNX checker and ONNX Runtime CPU inference:

- max absolute diff vs PyTorch: `0.0007368475`
- mean absolute diff vs PyTorch: `0.0000280040`
- mean relative diff vs PyTorch: `0.0053399284`

The 4-bit ONNX Runtime artifact replaced:

- `Gather`: `8` -> `GatherBlockQuantized: 8`
- `MatMul`: `9` -> `MatMulNBits: 9`

and passed ONNX Runtime CPU inference against the fp32 ONNX graph.

## Limitation

This is not a full in-browser TTS model yet. A complete browser version still
needs the autoregressive Higgs/Qwen3 text-to-codebook generator exported to an
ONNX Runtime Web compatible graph and a JavaScript sampling loop for the
8-codebook delayed-token generation.