trick to exactly adjust the volume of music for phase cancelling: - have the things lined up horizontally in audacity - find a bit of the mixed track where there's no talking, only background music - select the equivalent clip in the music track -

Effect

→

Amplify

, copy the default value in

Amplification (dB)

- ctrl+click on both of the header things on the left, switching the selection to the mixed track without moving the endpoints -

Effect

→

Amplify

again, paste the value into

Amplification (dB)

and copy

New Peak Amplitude (dB)

- select the entire music track -

Effect

→

Amplify

one last time, paste the new value into

Amplification (dB)

, hit

OK

you're basically subtracting the volume of the music track from the volume of the mixed track, giving you the exact amount you need to add to the music track to make it equal

mus + (mix - mus) = mus - mus + mix = mix

Hi, what is the best number of epochs to use on talknet? (I use the free version of google colab)

default

best is to look at the validation loss and wait until it drops to like > sub-0.14 for ridiculously large datasets (like 30 minutes or more), sub-0.10 for 150+ wav datasets, sub-0.09 for datasets between 100-30, sub-0.07 for 30 below (numbers courtesy of dawn)

are talknet melspecs really that similar to tacotron?

-oh talknet didn't notice

thankfully talknet has a thing where it'll tell you when it stops progressing

unthankfully you still need to reach the set epoch count before it'll let you continue

so when you notice the output being overtaken by

not in top 3

i recommend that you stop the cell, set the epoch count to a few more than you've reached, play it again, and then continue

Oh! Thank you! I have done several tests, but none of the results I like.

yeah talknet kinda sucks quality-wise

i gave it 2 hours of clean singing data and it still didn't match tacotron models with way less data

it sucks in general

how come

it sounds good to me

too picky

ಠಿ_ಠ Really? Some files I sent through actually deleted some sub harmonics of the voice.

It definitely depends on the voice and background noise

Yeah

would someone care to explain as to WHY Spamton's voice was removed

he sounded more legible and coherent than Jevil!

listen to this

what the heck is he even saying?

and THAT gets to stay on the website?

i know Spamton had lots of noise, but at least he was legible and understandable

need more proof? just watch (and listen) to this video i made

was completely unintelligible when i used him