Depends on the dataset parsing capabilities of the codebase you're usinf

Check to see how it splits the strings

And you can write a super quick python script to fix the dataset for you

I’m not sure I just downloaded this dataset from the MLP 4chan page. I know Tacotron takes a very specific format

Yeah I’m planning on parsing it

Some codebases might accept different dataset styles

Also I’m a little confused so the labeled file is sorted with all the speakers from the specific folder

Do I need to manually separate the speaker or can I just use the entire file as the transcript

Like if there’s Pinkie, Starlight, etc and all of the transcripts are in one file do I have to break the file apart?

You'd want a seperate transcript file for each speaker

Ok thanks!

Do all the filenames end in "!"?

Nah here's a few more examples:

29.792059    30.976211    00_00_30_Rainbow_Neutral_Very Noisy_Thanks big mac.
32.908777    33.450575    00_00_33_Rainbow_Happy_Very Noisy_Sweet!
33.692059    36.068303    00_00_34_Rainbow_Happy_Very Noisy_More cash to add to the stash!
39.335388    40.720990    00_00_39_Rarity_Neutral__How about a little break?

It's a really good dataset, quality is really high I've just never input data straight into tacotron so I was a little unsure how to handle it. But i guess the steps are to -> 1) parse file by removing the numbers in the beginning/formatting it. 2) Separate the speakers and make sure all speaker files are in one file

@User Modified Mega_bs script, you'd wanna add all the other speakers, maybe seperate them by the emotions too, but it could be something like. !mkdir wavs #@markdown # After running, put all wavs inside wavs folder. ra = open("Raritytranscript.txt", "x") rb = open("Rainbowtranscript.txt", "x") #@markdown # Begin renaming + Download zipped file. import os from google.colab import files transcriptRainbow="" transcriptRarity="" for count, filename in enumerate(os.listdir("wavs")): a= filename.Split("_") if "Rainbow" in filename transcriptRainbow= transcriptRainbow + "\n" + str(count+1) +".wav|" +a[6] if "Rarity" in filename transcriptRarity= transcriptRainbow + "\n" + str(count+1) +".wav|" +a[6] ra.write(transcriptRarity) #This should work? rb.write(transcriptRainbow) rb.close() ra.close() #Renames all files to right format, 1.wav, 2.wav etc. for count, filename in enumerate(os.listdir("wavs")): if "Rainbow" in filename dst = "Rainbow" + str(count + 1) + ".wav" if "Rarity" in filename dst= "Rarity" + str(count + 1) + ".wav" os.rename("wavs/"+ filename, "wavs/" + dst) !zip -r /content/output.zip /content/wavs/ files.download("/content/output.zip")

Oh thanks!

Np, think it should work or should be easy to make workable hopefully. I just found it a good excuse to practice/learn some python lol.

discord isnt showing the underscore but yea splitting by underscores makes sense to me.

hell yea!

You call this ultron?

Who is this imposter?

I was trying to make peter parker say JAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAMES!

(Metal gear meme)

Are ellipses discouraged (...) in transcribing? Do they affect the quality of the model at all?

the model isn't great at telling the difference between repeated letters. Ellipses can work fine for training but the model might get stuck if you use them for inference

Okay, perfect. Thank you!

Curious to know if this will get messier as I move up epochs

no, it didnt eeven reach 0.1