The Whisper autotranscriber really does work

(This dataset is of Buford from Phineas and Ferb)

Why waste hours transcribing text on your own when you can let an AI do it for you?

This gives you more time to do the things that you love.

Like maybe, building a rocket or fighting a mummy or climbing up the Eiffel Tower

(Gonna train Buford tomorrow because I’m being called to bed)

This gives me the idea of creating a speedrun category

Right?!?! Openai released a super good model

Perfect!

now i don't need to do it myself!

I’m working on a crust model how do I prevent it from going gibberish?

I have a dataset I’m gonna train with 18 seconds

And 10 wavs can someone help me out?

batch size 2, learning rate down to like 5e-4 maybe, and let it stew

Ok

Thanks I’ll get into it tomorrow

I ended up giving up I trained it wrong

how do you know what learning rate to use?

usually it's quite uniform

after x epochs of training use this lr, etc

well like the number letter combo is kinda confusing to me like do i do different letters for different things??

oh no

the letters are just to express smaller numbers

1e-3 vs 0.001

just stick to the defaults and it should be good

alr thanks that kinda cleara it up ig

I was thinking, is there any way to split the text for it to be synthesized on the Pipeline? Most of my models can speak longer phrases but I wanted so that any of the models are able to speak as much text as It can. I thought about splitting the phrases into chunks for it to be synthesized instead of generating one audio at a time and adding them together on an audio editor, does anyone know how to do that?

I have code to do this

I'll share with you tomorrow if you remind me