What about testig with anohter Datasets or Model?
Hi, I'm also currently researching to Increase and make faster model for Kazakh, Russian (mixed with Kazakh) languages
finetuning model distil-whisper/distil-large-v3 looks better, but needs som more data?
Now I'm using this model (distil-large-v3) in real time, but it's only for English. But also recognizes KK, RU and auto translates to English.
I tried the difference distil faster 2x, and result of this models is need to be one more pre trained.
I am currently working on collecting new data. Just human conversations on real live, and I will use STT yc and then validate the data with human work. After that we can try to tune the new model.
Have you tried to use ISSAI Kazakh corpus for PEFT?
Have you tried to use ISSAI Kazakh corpus for PEFT?
I tried, but its so slow, but i didnt tried to PEFT.
in this model WER is ok, but it only for clear speech.
in real data it is not so good. and using in conversation mixed kk+ru is also have less WER than whisper.
Write you TG