Qwen

/

Qwen3-ASR-1.7B

like 357

Follow

Qwen 67.8k

I was really excited to finally see this model be released publicly!

But upon testing it, forced alignment doesn't seem to be performing that well and outputs inaccurate timestamps a lot of the time, not making it that useful.
Is there any way to improve this without resorting to VAD?
[0] 'あれ': 46.560s -> 46.560s
[1] 'みんな': 46.560s -> 46.560s
[2] 'もう': 46.560s -> 46.560s
[3] '帰っ': 46.560s -> 46.560s
[4] 'た': 46.560s -> 46.560s