Horizontal g/total | AI HUB | Page 1

nova wasp May 31, 2025, 6:33 PM

#

Horizontal g/total with down spikes, is it normal behavior?

dark terrace May 31, 2025, 6:44 PM

#

dataset size and batch size?
the graph looks bad

fallow seal May 31, 2025, 6:46 PM

#

nova wasp Horizontal g/total with down spikes, is it normal behavior?

The spikes are RVC learning silence

dark terrace May 31, 2025, 6:52 PM

#

fallow seal The spikes are RVC learning silence

the mute file shouldnt appear that often in the graph, its bad

fallow seal May 31, 2025, 6:53 PM

#

dark terrace the mute file shouldnt appear that often in the graph, its bad

Watch him set the max amount of mutes misc_trolley

dark terrace May 31, 2025, 6:53 PM

#

misc_trolley

#

most probably the dataset is like 5 mins

nova wasp May 31, 2025, 6:58 PM

#

dataset 26 min and batch size 8

fallow seal May 31, 2025, 6:58 PM

#

Is the silence truncated?

nova wasp May 31, 2025, 7:00 PM

#

yes

dark terrace May 31, 2025, 7:01 PM

#

nova wasp yes

with this?

fallow seal May 31, 2025, 7:01 PM

#

Are you using the default amount of mutes?

nova wasp May 31, 2025, 7:02 PM

#

I used strip silence in Studio One but basically the same configuration

dark terrace May 31, 2025, 7:03 PM

#

uhm

#

well audacity it's safer since everyone uses it, no issues at all

nova wasp May 31, 2025, 7:04 PM

#

a piece of the dataset

dark terrace May 31, 2025, 7:05 PM

#

looks like the model is learning silence
maybe something went wrong during the preprocessing? how many slices you got?

nova wasp May 31, 2025, 7:05 PM

#

wait let me see

#

I put it in simple mode, got 584 slices

fallow seal May 31, 2025, 7:08 PM

#

Did you touch mutes at all?

nova wasp May 31, 2025, 7:08 PM

#

no

#

I trained another voice last week in the same way and everything went well

#

Do you think there's a lack of silence?

fallow seal May 31, 2025, 7:09 PM

#

nova wasp Do you think there's a lack of silence?

Those spikes are due to silence

dark terrace May 31, 2025, 7:10 PM

#

maybe there's too much silence

#

take the non-truncated version of the dataset

#

then truncate it using audacity, not studio one

dark terrace May 31, 2025, 7:11 PM

#

dark terrace with this?

using these settings

nova wasp May 31, 2025, 7:11 PM

#

okay

nova wasp May 31, 2025, 9:32 PM

#

did with audacity

#

#

#

with smoothing turned on it wasn't even falling before, now it is, but I still don't know if it's right

dark terrace May 31, 2025, 9:37 PM

#

nova wasp

ignore the loss g/total graph, is not precise
focus only in the loss avg50 one

nova wasp May 31, 2025, 9:37 PM

#

so it's good now?

dark terrace May 31, 2025, 9:38 PM

#

hmm seems to be less noisy than before

#

also use smoothing 0.5

#

for more accurate reading of the avg50 graph

nova wasp May 31, 2025, 9:39 PM

#

ok

#

I left a lot of breath in the dataset, maybe that's it?

dark terrace May 31, 2025, 9:40 PM

#

nova wasp I left a lot of breath in the dataset, maybe that's it?

breaths are good

#

model needs them to learn how to reproduce breathing

nova wasp May 31, 2025, 9:42 PM

#

I have a question, I'm using pretrained from the same language as this dataset. Could this influence anything? The vocal I trained last week was 15 minutes long

#

Maybe because this is almost 30 minutes long

dark terrace May 31, 2025, 9:45 PM

#

bigger datasets will be always better than smaller datasets

#

graphs don't need to be extremely smooth either

#

something in between

nova wasp May 31, 2025, 9:46 PM

#

Great then, thanks for your help

#Horizontal g/total