@User I wont say its perfect. She does do a bit of freestyling at the end there lol. Just too long a prompt for her.

This is really good

For a goodnight sleep

how many epochs did you do for yours, mine also has lower than 100 audio clips but when i make mine it is complete gibberish most the time?

I couldn't tell you unfortunately. I just waited till loss was below 0.15 and the graph was diagonal. It was the first model I made, 51 audio clips, longish sentences.

But yea I definetely think it could be due to how clear she is. I also made sure not to put any sort of punctation other than full stops

i thought you needed to have commas or something to show pauses in the audio? at least thats what i always do when making datasets

I just decided to keep it simpler. I dont use commas in transcripts or when putting in prompts for inference.

Share your dataset?

Ok I would definetely make your sentences shorter. Atleast any time you use a ? or full stop, seperate them.

I also get rid of apostrophes, could just find and replace them.

is there any reason that you remove apostrophes or just to make it simpler?

I just get the impression that it might trip the model up or not recognise them? Then with one of the first models I tried, it worked a lot better without punctuation.

This flippity dippity-hippity hip-hop You don't really wanna get into a pissin' match With this rappity brat, packin' a MAC In the back of the Ac' Backpack rap crap, yap-yap, yackety-yack

If the model can utter these lines

It's ready

But yea definetely seperate your audio sentence by sentence. I also seperate short words if they have a lot of silence before the speaker continues. The longer each wav the more attention required which I think would just trip the model up and generate gibberish.

i wierdly always thought the opposite, i thought the AI would do better with longer because theres more to train off of. but, thanks for the tips!

So I might be very wrong

But the way I would assume is that with these models shorter audio files = the model better adds that particular word to its vocab. Longer sentences = more natural flowing sentences. Super long sentences to me would require the user typing in something very similar for it to produce good results.

i do two or more sentences so the ai knows for how long it must pause between sentences

Ah fairplay, your models are pretty good

Yea I could be way off then.

That does make sense actually

my dataset transcripts range from 1 to 2 sentences somtimes 3

How do you deal with punctuation? Apostrophes and commas and the likes?

for compound words i separate them with spaces rather than hyphens (so that cmudict converts words properly) and for places where an ellipsis would be the good i just use commas instead

i try to avoid repeated punctuation

Ah kk, what about apostrophes?

i dont use apostrophes that much but i believe think merged dict has entries for those