Qwen3-TTS-Tokenizer-12Hz-ONNX-WebGPU
The speech decoder of Qwen/Qwen3-TTS-Tokenizer-12Hz (the codec every Qwen3-TTS 12 Hz model shares), as ONNX with quantized weights for running in a web browser with ONNX Runtime Web's WebGPU backend, as used by Quick Inference's Qwen3-TTS models.
Changes from the original
Exported by tools/export/export_qwen3_tts.py in Quick Inference with a
dynamic number of frames (the float export matches the PyTorch decoder to 4e-5), then quantized without calibration
(round to nearest, symmetric, blocks of 32):
tok_decoder_q4.onnx(96 MB; 457 MB in float32): matrix multiplications to 4-bitMatMulNBits, convolution weights to blocked int4DequantizeLinear(opset 21).tok_decoder_q4c8.onnx(124 MB): the same with int8 convolution weights (no brightening of the high band).
These are modified files: their output is close to, but not identical to, the original decoder's. Inside the Qwen3-TTS 0.6B CustomVoice pipeline, Whisper small.en's word error rate on eight test sentences was unchanged.
Use
codes int64 [batch, 16, frames] → audio float32 [batch, 1, frames × 1920], 24 kHz, clamped to [-1, 1]. The
decoder is causal (causal convolutions, attention over 72 frames), so long utterances can be decoded as the original
does: chunks of 300 frames with 25 frames of left context, dropping each later chunk's first 25 × 1920 samples.
Licence
Apache 2.0, as Qwen3-TTS (LICENSE).