phoonnx-qwen3-tts

ONNX export of Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice for phoonnx. The weights, the architecture and the nine timbres are the work of the Alibaba Qwen team; this repository only holds the graphs in ONNX form, converted with the script in scripts/conversion/qwen3tts/.

Graphs

File	What it does
talker.onnx	28-layer talker, one KV-cached step per 80 ms frame
text_embed.onnx	text ids to projected text hidden states
codec_embed.onnx	talker codec ids to hidden states
code_predictor_prefill.onnx	talker hidden plus code group 0, gives group 1
code_predictor_step.onnx	code group n gives code group n+1
sub_codec_embed.onnx	code-group token to its group's embedding
codec_decoder.onnx	16 code groups per frame to 24 kHz audio
tokenizer.json	the model's own Qwen2 subword BPE

All graphs are float32, opset 17 (18 for the codec decoder).

Parity

Checked against the upstream PyTorch model on CPU with greedy decoding:

Measurement	Result
Prompt embeddings, max abs diff	2.4e-07
Talker prefill logits, max abs diff	2.3e-05
Talker decode logits, max abs diff	5.0e-05
Greedy token agreement, 52 frames x 16 groups	100 % (832/832)
Codec decoder waveform, max abs diff	9.9e-07

Voices

Nine timbres: vivian, serena, uncle_fu, dylan (Beijing), eric (Sichuan), ryan, aiden, ono_anna, sohee. Ten languages: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian.

Use

from phoonnx.model_manager import TTSModelManager

manager = TTSModelManager()
manager.merge_default_voices()
voice = manager.voices["qwen3tts/ryan/en"].load()
for chunk in voice.synthesize("The quick brown fox jumps over the lazy dog."):
    ...  # chunk.audio_float_array

Licence

Apache-2.0, the licence of the original model. Cite the Qwen3-TTS technical report (arXiv 2601.15621) when you use these weights.