raven oriole
🎙️Model Link: https://huggingface.co/MatiasCoversAI/MusicAIBySingers/resolve/main/TheWeeknd400e.zip
Model Created By @raven oriole
🎤Dataset lenght: 16:00
🎤Sample Rate: 32K
🎤Batch size: 8
🎤Pretrain: KLM 4 HIFIGAN
🎤Epochs 400
🎶Trained using: All 7 songs on the album
Treined on Applio Colab
Sample (FP32):
grave impBOT
grave impBOT
mighty osprey
very low quality
Yeah, it sounds clean but the quality is weird.
This can tend to happen if you commit the mistake of including muddy vocals and/or overtrain/undertrain the model.
raven oriole
very low quality
Exactly should I try to train him at more epochs do you recommend?
Why the dataset is clean but if you want I can train it to more epochs so that it sounds of better quality
And the dataset had a lot of choruses and a lot of reverb and I removed it and I did it with the dataset cleaning models and that's what I did, I edited the audio, everything was worse, I think what it would need is more training.
And use the AI Hub guide to do the dataset process.
quick cedar
And use the AI Hub guide to do the dataset process.
i would say use the best vocal extraction models, dont overclean ur data, use tensorboard to get the best stepcount, and train at the right sample rate
u trained at 48k when ur dataset is under 32k which is why its so lq
by that i bet u use uvr deecho dereverb and ripped the songs off yt, uvr deecho ruins the dataset and takes off a ton of fidelity
raven oriole
i would say use the best vocal extraction models, dont overclean ur data, use te...
Hey Bro, look at the songs, I didn't get them from YouTube. I downloaded them in high-quality Hi-Res from the studio. I analyzed the audio with spek and it was fine and it's weird and it doesn't aggressively clean up the audio of The Weeknd's voice And I didn't copy the songs from YouTube, I downloaded them just like I told you before. And I did that training at 32k because that was what my dataset recommended for me to do in that training. Everything you just explained to me is as the model says but I did it in 32k also because it was recommended to me. Lyery
raven oriole
u trained at 48k when ur dataset is under 32k which is why its so lq
What uvr5 models do you use to clean your datasets?
raven oriole
by that i bet u use uvr deecho dereverb and ripped the songs off yt, uvr deecho ...
I ask you To be able to clean the songs
Returned
So I can make the model back and update it here for more periods and better audio quality for the weekend model.
quick cedar
Hey Bro, look at the songs, I didn't get them from YouTube. I downloaded them in...
ur data is 32k but u trained at 48
raven oriole
Yes, but I didn't take the songs off YouTube xd
quick cedar
Yes, but I didn't take the songs off YouTube xd
oh well that means uvr deecho took ur sample rate down by a lot
u should use melroformer to extract the vocals and melroformer to dereverb
raven oriole
ur data is 32k but u trained at 48
I trained the model at 32k xd
quick cedar
I trained the model at 32k xd
oh thats odd then, its supposed to cut off at the sample rate u train at
raven oriole
u should use melroformer to extract the vocals and melroformer to dereverb
Ok, and what are those vocal melroformer and dereverb models called?
quick cedar
Ok, and what are those vocal melroformer and dereverb models called?
theres a lot of models and the best depends on what u prioritise, u can try a lot of em out on mvsep.com
make sure u login or the line will take forever
raven oriole
oh thats odd then, its supposed to cut off at the sample rate u train at
Of course, that's why it seems strange to me.
quick cedar
Of course, that's why it seems strange to me.
i would just use melroformer, reclean ur dataset, make sure u dont overclean it and retrain
should fix problems
raven oriole
theres a lot of models and the best depends on what u prioritise, u can try a lo...
Okay bro, I'm going to try to clean up the songs to separate the vocals and separate the reverb in mvsep
raven oriole
i would just use melroformer, reclean ur dataset, make sure u dont overclean it ...
Ok, yes, that's what I'll do with the mvsep to clean the songs better, the vocals, remove the dereverb. Do you recommend that when you have cleaned the songs, I edit the vocals? And what model do you recommend to remove noise?
And choirs or double voice
quick cedar
Ok, yes, that's what I'll do with the mvsep to clean the songs better, the vocal...
for noise use spectral denoise in rx11
quick cedar
And choirs or double voice
mel roformer kareoke
raven oriole
for noise use spectral denoise in rx11
Ok I'll use the rx11
raven oriole
mel roformer kareoke
Well I'll use that one too. 👍
High quality songs weigh almost 80 mb xD
raven oriole
imma try to make a weeknd dataset too
Okay bro, that's very good. 👍
raven oriole
That is, I only used 7 songs from the album, that would be it.
raven oriole
what does your post mean by "all 7 songs"
Now I'm cleaning the songs I'm removing the reverb.
quick cedar
Now I'm cleaning the songs I'm removing the reverb.
btw none of the vocals should have noise
so dw about the spectral denoise
raven oriole
so dw about the spectral denoise
And how can I remove the noise once I've removed the reverb?
raven oriole
btw none of the vocals should have noise
Send me the audio to DM
quick cedar
And how can I remove the noise once I've removed the reverb?
should be no noise, he records in a pretty noiseless studio
raven oriole
Give them then perfect 👍
quick cedar
Give them then perfect 👍
sent the thingie
raven oriole
sent the thingie
If you send me that audio privately, please send it to me here.
quick cedar
If you send me that audio privately, please send it to me here.
cant bc razer gets mad at copyright shi
raven oriole
cant bc razer gets mad at copyright shi
Yes, don't send it to me here, send it to me in DM, it would be private because here it is deleted, you don't have to send it because of copyright
quick cedar
Yes, don't send it to me here, send it to me in DM, it would be private because ...
i did already
raven oriole
i did already
Ok
wooden wharf
lol
raven oriole
xd