[examples] better results on wenetspeech using revised transcripts

[xingchensong]

⚠️quality is important than quantity⚠️

see wenet-e2e/WenetSpeech#54 for more infos

Model	Dev	Test_Net	Test_Meeting	blankpenalty
u2pp_conformer (original transcripts)	8.37 % N=328207 C=302993 S=12556 D=12658 I=2253	9.02 % N=414285 C=379271 S=23881 D=11133 I=2365	15.52 % N=220358 C=187746 S=14507 D=18105 I=1596	0.0
-	7.10 % N=328207 C=308457 S=13211 D=6539 I=3537	8.83 % N=414285 C=381933 S=24810 D=7542 I=4215	13.63 % N=220358 C=194446 S=16232 D=9680 I=4132	2.5
u2pp_conformer (revised transcripts)	5.90 % N=328207 C=311721 S=14597 D=1889 I=2888	8.96 % N=414092 C=380232 S=27606 D=6254 I=3222	11.99 % N=220358 C=195808 S=18243 D=6307 I=1878	0.0
zipformer-M	8.58	7.84	14.64	0.0
-	7.24	7.77	12.37	2.5

[xingchensong]

update whisper result:

Model	Dev	Test_Net	Test_Meeting
whisper (original transcripts)	7.27 % N=328207 C=308016 S=11392 D=8799 I=3672	7.90 % N=414097 C=383382 S=18954 D=11761 I=2018	13.00 % N=220358 C=194417 S=11788 D=14153 I=2705
whisper (revised transcripts)	4.66 % N=328207 C=315591 S=10352 D=2264 I=2692	6.54 % N=414092 C=389523 S=19101 D=5468 I=2513	8.84 % N=220358 C=202722 S=11296 D=6340 I=1839
paraformer	3.38 % N=328207 C=319078 S=8045 D=1084 I=1959	6.74 % N=414285 C=388817 S=20119 D=5349 I=2444	6.95 % N=220358 C=206461 S=9090 D=4807 I=1422

As you can see, the overall results are already very close to those of Paraformer. I believe that by adding more open-source data, Whisper can definitely achieve better performance than Paraformer.

[Ryuk17]

Any plan to release this trained model?

[xingchensong]

Any plan to release this trained model?

目前效果还没全面超越paraformer，没有release的意义，另外听闻阿里有意发布更好的多语种模型，后续在那个更强的base模型上去调可能会更好

[xingchensong]

@Ryuk17 另外，现在在wenet训练200M级别的模型已经很高效了，8卡3090，使用deepspeed 3~4天就能跑出来上面的u2pp_conformer结果

[Ryuk17]

了解，谢谢

[xingchensong]

Any plan to release this trained model?

传了一个whipser的

链接：https://pan.baidu.com/s/1R1mJyqc_MwKZy6Vn32Z3zA?pwd=171n
提取码：171n
--来自百度网盘超级会员V6的分享

[wwfcnu]

Any plan to release this trained model?

传了一个whipser的

链接：https://pan.baidu.com/s/1R1mJyqc_MwKZy6Vn32Z3zA?pwd=171n 提取码：171n --来自百度网盘超级会员V6的分享

这个微调后的模型能重新转回到openai style的ckpt吗

[Zth9730]

Any plan to release this trained model?

传了一个whipser的
链接：https://pan.baidu.com/s/1R1mJyqc_MwKZy6Vn32Z3zA?pwd=171n 提取码：171n --来自百度网盘超级会员V6的分享

这个微调后的模型能重新转回到openai style的ckpt吗

https://github.com/wenet-e2e/wenet/blob/main/wenet/whisper/convert_whisper_to_wenet_config_and_ckpt.py#L199 参考这里是可以转回去的