| --- |
| license: apache-2.0 |
| base_model: Qwen/Qwen3-ASR-1.7B |
| pipeline_tag: automatic-speech-recognition |
| language: [en, zh, yue] |
| tags: [qwen3-asr, medical-command, native-windows, streaming] |
| --- |
| # PhonoCommand ASR |
| Current Web V15 weights based on Qwen3-ASR-1.7B. The R17 pipeline selects between |
| V15 and the pinned original model using acoustic duration and language evidence. |
| It uses accumulated-audio streaming, conservative text-prefix reuse and one |
| generation per update/final. No command keyword rewriting or chat rejection. |
|
|
| Read [ALGORITHM.txt](ALGORITHM.txt) and [WINDOWS_HANDOFF.txt](WINDOWS_HANDOFF.txt). |
| [reference_code.zip](reference_code.zip) contains minimal native PyTorch/SDPA |
| inference and acoustic-policy references, not a Web application or full scheduler. |
| The full upstream ASR SDK and vLLM are not required by this reference loader. |
|
|
| Use [sync_models.py](sync_models.py) to download the paired release and pinned |
| dependencies. [release_set.json](release_set.json) pins both component commits; |
| MANIFEST.json hashes every artifact. No recordings/training rows are published. |
|
|
| Native implementation validated on the Linux build host: priority recording |
| 16/16 final reference windows; Windows hardware remains to be tested. |
| Partial hypotheses may still be wrong. Timing is measured from VAD end, not |
| manually annotated word ends. See VALIDATION.json. Existing Linux vLLM memory |
| and latency measurements do not establish Windows performance. |
|
|
| Apache-2.0 Qwen derivatives; Silero helper retains its separate MIT license. |
| SenseVoice is referenced from its original repo under its original license and |
| is not redistributed. Language helper evidence is probabilities only. |
|
|
