what an improvement

s t o n k s

lol

That's great alignment already. If that's from the improved Mr. Beast model, I think its ready to be submitted.

as anyone made a model of a conventional tts voice yet

I feel like that would be silly but also very easy and maybe fun

How many samples should I give a model for like a clear voice

Thx

Epoch 68

Hey I'm working on building a web app to help people transcribe characters from tv shows if you have the episode and the subtitles. My issue was I have the subtitles but they weren't matched up correctly and they weren't labeled by character. I want to get Rick Morty Beth and Jerry and Summer before the 20th but this should speed things up. but Im building it so its very general that you can do any tv show you have the subtitles and audio. my concern is that all my transcription clips are gonna be only 1-4 seconds i have an idea of how i would solve that but wondering if yall think that would be a big issue

less than 1.5 sec clips actively hurt the model

dam i really didnt want to have to fix that problem

I'd just discard the short ones

@User ayo any help in #841732367962210344 ?

how are you doing speaker segmentation?

pydub

I am using a similar thing

how that exactly?

I mean its not that good im testing it out rn

and by the thing you sent there seem to be empty lines

this is gonna take a while

even with 50 batch size

fuck

out of memory

lemme try 30

i get the timestamps from the subtitles file and then just clip that start and end from the wav. The thing is that the subtitles aren't accurate to the time so you need to correct them with an offset and sometimes they are wrong which is why i made this so that I can get good data. also the subtitles aren't the whole sentence sometimes if the sentence is too long so you could have one character talking for multiple subtitles (Honestly I was thinking that if you just had some labeled you could just finetune a Speech2Text model from hugging face to do the rest)

or just make your own model that can label characters and write the transcript. then you could just feed who seasons into it and get nice voice data

YOOOO

ITS LOOKING GREAT OFF THE BAT