FireRedASR icon indicating copy to clipboard operation
FireRedASR copied to clipboard

请问对粤语的支持怎么样?

Open zhangchongcool opened this issue 10 months ago • 5 comments

zhangchongcool avatar Feb 21 '25 02:02 zhangchongcool

没具体测过。但是模型应该是有一些粤语歌曲+粤语能力的。有兴趣可以找个测试集测一下,能反馈结果的话~提前感谢

FireRedTeam avatar Feb 21 '25 03:02 FireRedTeam

It is very bad. Tested on random sample of 200 items from Common Voice 17 Validated zh-yue, CER is basically 100%

Attached detailed results, TSV (wav file, target, hypothesis, RTF using RTX4090)

asr_cer_firered.txt

kmn1024 avatar Mar 05 '25 05:03 kmn1024

Me also got worst results on 8 test sets, the cer is around 75~90%. However, the result of Sichuan accent is way more better then Cantonese.

AdilAdam avatar Mar 10 '25 02:03 AdilAdam

没具体测过。但是模型应该是有一些粤语歌曲+粤语能力的。有兴趣可以找个测试集测一下,能反馈结果的话~提前感谢

是否有计划开源一下lora微调的脚本,能用common voice的粤语子集进行微调

fuxiao-zhang avatar Mar 18 '25 10:03 fuxiao-zhang

效果很差,我用了粤语天气预报和新闻播报试了,音频中无背景音干扰,真的不太行。

OnlyFlashEobard avatar Apr 02 '25 06:04 OnlyFlashEobard