Traning on wenetspeech couldn‘t converge

[dyyoungg]

I compared two experimental data setups.
setting 1: WenetSpeech（Chinese）only
setting 2: Wenet + Giga （about 1：1， Chinese + English）

It's interesting that training on setting 1 can't decrease normally （blue curve in the following image）， while setting 2 mixed with English can converge normally.
Have you observed this phenomenon in your experiments?

[jishengpeng]

I compared two experimental data setups. setting 1: WenetSpeech（Chinese）only setting 2: Wenet + Giga （about 1：1， Chinese + English）

It's interesting that training on setting 1 can't decrease normally （blue curve in the following image）， while setting 2 mixed with English can converge normally. Have you observed this phenomenon in your experiments?

This situation is somewhat unusual. You may use a small amount of Chinese data (approximately 500 hours) to verify whether this issue always arises when the model is trained on purely Chinese data.

[wntg]

I‘m intersting in Chinese too. Do you have any further results?

[boltzmann-Li]

WenetSpeech could be too noisy, you may want to start with AIShell3, then WenetSpeech4TTS.