with my

half as interesting

model i found a specific phrase was spoken more flatly in a later version of the model

I kept trying to tell them that they should just stop the training when it gets to the right val loss but they seem insistent on training to a huge amount of epochs

Also what is the logic they're using to decide that they should train to the number of epochs that's double the amount of their wav files

does anyone have a good site that can remove background music? I see a couple online, but was wondering if someone has something they always use

mvsep’s mdx b option works well

no wait was it mvsep yes its mvsep

sometimes audo.ai works pretty damn well sometimes it just, doesnt do anything at all for some reason

dang, i really want to grab like every site to do tests to see which ones works well

well mvsep has a few options

mdx a and b, spleeter, some other stuff

yea, hmmm. I'm going to need to read up the difference between these

have you tried facebook's isolator?

!pip -q install demucs
!demucs -n mdx_extra "audio.wav"

oh, interesting

oo, wait, i just found a youtube video of someone comparing spleeter, demucs, and lalala. HA, now I don't need to do any testing

dang, except they're all not perfect

mdx b from mvsep seems to be the best so far. LALAL cost money. Media.io is free but it's bad results

The notebook one I've tried. The recommended settings seem 2 b 2 aggressive for some but otherwise do good

for me it's not aggressive enough, leaves a ton of music in

though oddly enough i once fed it a mostly-phase·cancelled track and while it sounded bad normally it suddenly became better when i mixed it down to mono

Is It Possible To Train Talknet On Android?

the training happens in a browser so it is possible, albeit inconvenient you'll have to have the page open in desktop mode (else it might not load the code), keep it in front while it's training (i recommend splitscreen if your phone supports it), and also go into your google drive every so often to clean the trash that accumulates during talknet training

Aaaah Ok

Thx

ooh

someone actually already did the character sorting for puyo puyo tetris characters

i eventually wanted to turn that into dataset form but haven't gotten around to it yet

You should try audo.ai

way better

if you can find the music you can try putting it in audacity and syncing it up exactly to the video and inverting it and adjusting the volume (there's a trick for that) and then doing ai vocal extraction