It isn't

I wouldn't go over 200 epochs in any case

I'll check it out soon

40-80 mins is a good range

and mostly stay way below that

it's not baking a cake

I did go lower than that and tested it on like 10, 20 and 30 minutes. It barely talked

Anyways, since you're awake now, can you give me an example of how do I add Arpabet into the transcript text file? Because my method of doing "wavs/1.wav|{H EH1 L OW1} {W ER1 L D}." doesn't seem to work well for me... Try to find stuff on Google leads to almost no guides at all...

I was awake the whole time lol

Oh...

what did you do?

did you train the model on arpabet input?

"I'm testing the model for speaking normal and all, I think it's alright but not like 100% perfect" -40 mins training

it a multivariate regression problem how long to train

ye

samples and phonemes

it's not as easy as baking a cake at a certain temp for a certain time

Well, I did like some other datasets I researched where the normal text is placed first then other type of input, to which case I added the way the Syntheizer colab notebook outputs the Arpabet, being separated by another "|" in the middle.. For my Mei model, it worked with the normal raw_input = True texts, but for Arpabet, just about 25%-50% working with some phonemes...

You can really just play around with words

like my dream model said badboyhalo where like the halo part had a weird a

I changed it to bad boy hay low

and bam

I know, but I'm just like experimenting around so I could make an Arpabet enabled model in the future lol

yeah sometimes you gotta screw around with the words to make it say stuff right

So yeah, technically I did train it on Arpabet, but in the wrong way

you can only expect it to work if you trained on arpabet input

and you can't mix arpabet and normal input (!)

(someone did that at some point and asked me why it didn't work xD)

Could you just make a new line in your transcript, and make the same audio file have another transcript but in ARPAbet?

the whole transcript you train on needs to be either arpabet OR normal text