Discrepancy in Qwen3-ASR WER results on wenetspeechyue datasets compared to official paper
Hi thank you for your great open-source work!
I noticed that the Word Error Rate (WER) metrics reported in this repository for wenetspeechyue_long (11.66) and wenetspeechyue_short (17.77) show a significant discrepancy compared to the official results from the Qwen3-ASR paper (which are 8.85 and 5.82 respectively).
I conducted some local evaluations on my end, and my test results (9.42 for long and 5.98 for short) are much closer to the data reported in the Qwen3-ASR paper.
Could you please share more details on how the evaluation was conducted in this repository? Thanks again for your time and contribution!
Hi there
Thanks for pointing out. So far all the model listed in the evaluation table were run without forcing language id, we will update the table with more completed figures soon.