Qwen3-ASR 0.6B — ONNX for WebGPU
Multilingual speech recognition (used for Chinese) that runs in the browser with onnxruntime-web on WebGPU. It is used by KamVox (Transcribe), where audio never leaves the user's device.
This is a converted copy of Qwen/Qwen3-ASR-0.6B.
What was changed (Apache-2.0 §4b notice)
The original weights were modified by KamVox:
- exported from PyTorch to ONNX (an audio encoder graph and one fused decoder graph with a KV cache);
- quantized to int8 (MatMulNBits, block size 32) for the encoder and decoder, and int8 with fp16 scales for the token
embedding table (
embed_tokens.int8.bin, looked up outside the graphs); - the tokenizer was reduced to a decode-only table (
tokenizer.json, custom format).
Quality: on 30 FLEURS Mandarin clips the int8 model gives a character error rate of 0.19 % against the original PyTorch
model (28 of 30 identical). The graph contract, export scripts and parity checks are in the KamVox repository
(docs/asr-graph-contract.md, tools/asr-export/).
Files
| File | Purpose |
|---|---|
encoder.onnx |
audio encoder (int8) |
decoder.onnx + decoder.onnx.data |
text decoder with KV cache (int8) |
embed_tokens.int8.bin, embed_tokens.scales.f16.bin |
token embedding table |
tokenizer.json |
decode-only vocabulary |
config.json |
model, mel and prompt settings |
manifest.json |
file list with sizes and SHA-256 |
Credits and license
- Base model: Qwen3-ASR-0.6B by the Qwen team, Alibaba Cloud — Apache-2.0.
- ONNX conversion builds on ideas from andrewleech/qwen3-asr-onnx and OpenVoiceOS/qwen3-asr-0.6b-onnx (Apache-2.0).
This converted model is distributed under the Apache License 2.0 (see LICENSE).