3 wavs? Would shoot towards having less than 0.10 loss value to start

This is obviously an improvement over sounding like LJSpeech, but it's not right on the mark.

Okay, but where did I go wrong was in the epoch?

Nothing wrong with the epoch, it's mostly the loss you need to watch for. Some models may need more epochs to train on, some may not

Oh, I understand

but it took me almost 1 hour to wait for the model to be ready

I've only gotten accidentally lucky twice on each nb so far.. 😅 Consistent testing is definitely avoids any risk any any further pains of trying to correct models

The lucky ones weren't as great as they probably could've really been either imo

Later I will see how epoch works

Speaking of that here is a sample from one of that lucky 1 wave attempt.

Can't dl audio from the first model

The voice sounds so cool

no robotic and no repetitive

I did start training it with a very low LR and kinda fell asleep halfway through so I'm actually surprised it turned out good. Sample using the earliest model i can grab audio from which does use more wavs (even more if I can get my shit together and do it :))

that's nice. you guys from uberduck create a lot of perfect voices

What does HIFI-GAN actaully do?

I know what the other 2 steps do, but not Hifigan.

I'm talking in relation to Talknet2 model training

basically, talknet doesn't output raw audio because that's too hard to measure the quality of instead it outputs something called a spectrogram, which is a 2d image of what frequencies play where and then it's the job of hifi-gan to turn that into actual audio

Ahh

So why does running it longer make it sound better?

Also, I have a singer who has a somewhat rigid range of singing. Because I have not found any data of him singing certain notes, the AI singing will yell the sounds of a seizure in place of those notes. But once I give it a note I've actaully given it, it will do well. Is there a way to artificially generate these spaces in the vocal range? Is there a bit of the training I might be able to train longer to generate this?

because you're training it to get good at specifically converting the current model's output into stuff that sounds like the training data the base model is already pretty good granted

you could try making pitch-shifted copies of the data using something like melodyne or newtone or vocalshifter

Would that be cheating or do you just need to do that sometimes?

why would it be cheating

why would anything be cheating really

I'll start keeping a note of every half-step I have musical data of

if it gives a better result then by all means do it