pip install -r requirements.txt

# FP32
python generate_onnx.py --text "Give every small business the voice of a big one." \
  --instruct "Speak in a warm and friendly female voice" \
  -o output_fp32.wav

# INT4 (~4x smaller, faster)
python generate_onnx.py --variant int4 \
  --text "Give every small business the voice of a big one." \
  --instruct "Speak in a warm and friendly female voice" \
  -o output_int4.wav

# Chinese
python generate_onnx.py --variant int4 --lang chinese \
  --text "让每一家小企业，都拥有大企业的声音。" \
  --instruct "Speak in a warm and professional female voice" \
  -o output_zh.wav

Model Architecture

Qwen3-TTS is a three-stage autoregressive pipeline:

Text --> [Tokenizer + Embedding Construction] --> inputs_embeds
             |
             v
     [Talker LM]           28 layers, 2048 hidden
     predicts codebook group 0
             |
             v
     [Code Predictor]      5 layers, 1024 hidden
     predicts groups 1-15
             |
             v
     [Vocoder]             single forward pass
     16 codebook groups --> 24kHz waveform

The pipeline is split into 4 ONNX models:

Model	Description	FP32 Size	INT4 Size
`talker_prefill.onnx`	Full sequence prefill with KV cache output	5.3 GB	1.4 GB
`talker_decode.onnx`	Single-step decode with KV cache	5.3 GB	1.4 GB
`code_predictor.onnx`	Predict codebook groups 1-15	440 MB	322 MB
`vocoder.onnx`	Codes to 24kHz waveform	876 MB	558 MB

Repository Structure

.
├── config.json              # Model config (dimensions, token IDs, language map)
├── tokenizer/               # Text tokenizer (vocab, merges, config)
├── embeddings/              # Pre-extracted embedding weights (.npy)
├── fp32/                    # FP32 ONNX models
│   ├── talker_prefill.onnx
│   ├── talker_decode.onnx
│   ├── code_predictor.onnx
│   └── vocoder.onnx
├── int4/                    # INT4 weight-only quantized models
│   ├── talker_prefill.onnx
│   ├── talker_decode.onnx
│   ├── code_predictor.onnx
│   └── vocoder.onnx
├── generate_onnx.py         # Reference ONNX-only inference script
└── requirements.txt         # Inference dependencies

Supported Languages

English, Chinese, Japanese, Korean, German, French, Spanish, Italian, Portuguese, Russian.

Reproducing the Export

The export scripts are in the wavekat-tts repository:

bash

cd tools/qwen3-tts-onnx
pip install -r requirements.txt

# Export FP32, validate, and quantize INT4
make all

About WaveKat

WaveKat builds open-source voice pipeline components in Rust. This ONNX export is maintained as part of wavekat-tts, which provides unified TTS inference across multiple backends.

Acknowledgements

Qwen3-TTS by the Qwen team at Alibaba Cloud

wavekat/Qwen3-TTS-1.7B-VoiceDesign-ONNX

Qwen3-TTS 1.7B VoiceDesign (ONNX)

Quick Start