Fix eot

[Qiaochu-Song]

There is a bug when generating train.yaml from whisper checkpoint. Eot was using the token index of sot, and this pr fixes that.

[xingchensong]

good catch! How did you find it? Does it affect trainining results?

[Qiaochu-Song]

good catch! How did you find it? Does it affect trainining results?

I was trying to convert and load whisper into wenet, then tried some samples without any fine-tuning. I was expecting it to give some reasonable outputs. However it did not stop decoding, even after eot was produced. Other than that the transcription looks correct. After changing eot to the correct idx, wenet decoding functions is able to terminate decoding correctly.

Regarding effects on finetuning, I have not test it yet, but I assume it might affect training results a bit, but should not be too much, as tokens after won't be penalized in training.

[xingchensong]

I see, THX !!!

[xingchensong]

I can confirm that this fix does not affect training result.

• green: eot is 50258
• red: eot is 50257