Feature Request: Support for Qwen3-ASR

[hlevring]

https://github.com/QwenLM/Qwen3-ASR

[Nishant-ZFYII]

.take

[github-actions]

Thank you for looking into this issue! Please let us know if you have any questions or require any help.

[hlevring]

Honestly I am beginning to think its not worth it. When I made the request I had gotten some amazing results, but I was testing the wrong model (sounds a little stupid, and it its).

This demo here gives insanely good results - works like a dream.
https://huggingface.co/spaces/Qwen/Qwen3-ASR-Demo

It seems to be using another model called Qwen3-ASR-Flash (remote via API key) which is not open sourced

Qwen3-ASR-1.7B and 0.6B on the other hand seems to be a very different models. I tried to run their official demo https://huggingface.co/spaces/Qwen/Qwen3-ASR on some of my English sample audio files, and its making subtle mistakes including repetitions that I would not have expected.

For example:

“that’s that’s quite a quite exciting feature”
“would would be would be applied”
“which which works pretty amazing”
“we’ll we’ll mainly focus”
“but but still”

Qwen3-ASR-Flash is mind blowing good, but Qwen3-ASR-1.7B works terribly on my test files, even Whisper Small.en gives much better results.

Potentially it might be a bug that is causing this, so I will report in their repro - I just need to get some audio files and actual test-results I can share with them.