Qwen3-ASR 0.6B — asr.js ONNX bundle

This repository contains the exact custom ONNX export and tokenizer assets used by the local @asrjs/speech-recognition Qwen3-ASR adapter. The upstream Qwen/Qwen3-ASR-0.6B repository is Apache-2.0 licensed and does not provide ONNX graphs at that revision. This bundle contains an FP32 audio encoder, FP16 decoder prefill/step graphs and their shared external data. It is an experimental asr.js export.

asrjs-manifest.json lists the exact files and SHA-256 digests. The real-weight four-demo matrix accepted 16 of 18 requested fixture flows; native streaming was unsupported. This does not certify broad multilingual accuracy, all backend/precision combinations, or physical microphones. It is not a general approved execution profile.

Use the local artifact adapter in @asrjs/speech-recognition; these files are not a generic Transformers.js ONNX export. Export lineage and current limitations are described in the Qwen porting notes and model matrix.