Hungarian training data

[hekkr000]

Today I've tested the v2 model vs the v1 and found that in a lot of cases the v2 one tends to just spit out this text: " Feliratok az Amara.org közösségétől". I think there might be a lot of invalid data scraped from this site: amara.org. I mostly tested the models with music, here are a few examples:

https://youtu.be/1BI54w6T_Uo
https://youtu.be/M5CwqYQNRcY

I think it would be beneficial to look this string up and clean the training data.
Otherwise, I've seen a large improvement in Hungarian, whenever the whisper was willing to create real transcripts.

[SilverGreen93]

Hi,
I randomly get the same message in Hungarian in kdenlive using Hungarian Large model

Don't know what the issue is, but seems quite random to get stuck like this.