has anyone come across a tool that automatically removes unsuitable audio from a tts training dataset?

Here are the following classes of unsuitable audio I'd like it to detect: 1. Audio has too much noise in it. 2. Audio's speech does not match transcription speech. 3. Speaker is speaking too quickly to be intelligibly understood. 4. Audio captures a different speaker than the labeled speaker. 5. Audio speech contains out-of-vocabulary words.

I'm about to build a solution to do this so if anyone could save me a month of work by telling me that'd be great, haha

>out of vocabulary words

would that be an issue???

@Gosmokeless28 apologies for the ping ™️ but is it fine if i used your spongebob dataset to train on diff-svc?

That depends: Which SpongeBob dataset did you use?

Cuz one of them was originally made by Speaking of AI, not me

Hi I wish to train a voice clone for steve jobs and I got a rtx 4090... I just wonder is this gpu good enough for training?

I assume so, but why do you need your own GPU? Are you going to train the model locally or something?

Oh gotcha I can use Colab!

4090 is probably like 5x better than colab

that being said it does take some effort and knowledge to set up local training

i tried to make the pipeline notebook not depend on colab but it hasn't been tested outside of it

and it does still depend on linux, specifically debian/ubuntu/etc

ops I do not have linux though 😢

,,though come to think of it the parts that depend on linux are mostly just the dataset loader which you probably won't need

i know @Justin trains talknet on windows, idk about tacotron though

that would be the recent one in #835647732453605376

It's easy to set up with Anaconda

But there are two recent ones in datasets. Can you specify which one?

so hi

when i was trying to fit data in the modle

random forest regressor

im getting this error

how i can deal with it

I would think it was, why wouldn't it be? I always thought that transcription engines like Kaldi check audio against a pre-defined dictionary. At the very least I've gotten OOV errors when using forced word alignment tools like Gentle: https://github.com/lowerquality/gentle Is your experience different? I'd love to know if so, thanks @(Dawn) Will Draw Fictional Women

On this topic, is there any solution for easily turning Colab notebooks into web-accessible servers with RESTful APIs? I had to spin up a JS server on a GPU from AWS' EC2 and call python3 as a child process in order to do some audio synthesis for the prod env of my app. It was super manual, ugh

i've seen some notebooks use ngrok and cloudflare and such but for production i highly recommend against this, you'd have to manually restart it every day even with pro+ you should be able to synthesize without a gpu though, last i heard uberduck synthesizes on cpu

oh thats awesome if thats true, ill look into it! thanks