Logo
Sign in
Sign up

TF2 Spy | Dennis Bateman_e16-GPT_e15-SoVITS

Created: March 20, 2024
TF2 Spy | Dennis Bateman_e16-GPT_e15-SoVITS

TLDR the SoVITS is 15e with GPT being 16e
GPT is DPO trained
<https://huggingface.co/Lusbert/GPT-SoVITS/resolve/main/TF2_spy-16e_GPT-15e_SoVITS.zip?download=true>

notes: (for people who wanna learn more stuff about my experience with GPT-SoVITS)
dataset size is 21 minutes
batch size 2 on GPT and SoVITS
SoVITS Learning Rate 0.4
15 epochs sovits took around 40 minutes
16 epochs gpt 27 minutes with DPO
saving freq 2

at first i trained GPT to 20 epochs
but last epoch saved is e18
prob cause my saving freq was 3
so ye be sure to look out ofr that as well
as for SoVITS model it was fine most likely since i trained to 15 epochs and not 20
so prob smth to do with if the modulo of epoch is 0 or not so like

[epoch count]mod[saving freq]
with mod being % in python
so in my case where i trained to 20 epochs with saving freq of 3
20%3 which is 2 aka non 0, so thats BAD :3
so just make sure the mod is always 0 so that you train to the full epoch count

i didnt like the model when i first trained without DPO so i retrained GPT model with DPO
it also went from using 6GB VRAM to using 8GB
this time it took 35 minutes

also apparently in SOME cases
the default pretrained GPT model is better, often where new words are said or the punctuation is different i'd say? but im not really sure

Additional Details:

Tags: No tags available

Download Link: https://huggingface.co/Lusbert/GPT-SoVITS/resolve/main/TF2_spy-16e_GPT-15e_SoVITS.zip?download=true Hosted on Hugging Face

Convert this model in EasyAIVoice

Sample deck

Pitch previews

Listen across generated pitch samples, then Run. Hold Ctrl or Cmd to select multiple pitches.

samples ready
female base 路 pitch 0 Generated female reference sample
Run
male
female
female-base generated sample male-base generated sample Enrichment pending: no scores or voice accords yet Ctrl/Cmd-click = select multiple pitches