# machine-learning
- r
Reclezon
05/25/2022, 12:13 AMOh damn an English folder - r
Reclezon
05/25/2022, 12:13 AMforgot aI was gonna do that after - u
{K EY1} (Kei)
05/25/2022, 12:16 AMif you want you could learn the phonetics and type them in directly might be a bit of a chore, though or you could probably just use a romaji to hiragana converter of some sort - u
{K EY1} (Kei)
05/25/2022, 12:17 AMbtw kana includes hiragana. - u
(Dawn) Will Draw Fictional Women
05/25/2022, 12:17 AM1 minute 40 secs - u
{K EY1} (Kei)
05/25/2022, 12:17 AMoh i could def do that - u
{K EY1} (Kei)
05/25/2022, 12:17 AMwell. probably. but it doesn't seem too long - u
(Dawn) Will Draw Fictional Women
05/25/2022, 12:19 AMzamn? - r
Reclezon
05/25/2022, 12:29 AMYeah idfk shit about kana or whatever. It's 71 wavs each 5 seconds or less so eh - r
Reclezon
05/25/2022, 12:29 AMProbably wouldn't be worth it without some longer wavs? - u
{K EY1} (Kei)
05/25/2022, 1:29 AMit should be fine, but non-english def needs more data than english i got decent results with about 14 minutes of audio - u
{K EY1} (Kei)
05/25/2022, 1:38 AMthis is a model trained on about 14 min, hifigan finetuned 彼は私に自分の魅力をアピールしてきたけれど、タイプでは無かったので気づかない振りをした。 (kare wa watashi ni jibun no miryoku o apiiru shite kita keredo, taipu dewa nakatta node kizukanai furioshita.) - u
(Dawn) Will Draw Fictional Women
05/25/2022, 1:43 AMif ya want wavs are in #835647732453605376 - u
(Dawn) Will Draw Fictional Women
05/25/2022, 1:43 AMi have pro so i can train if i figure out that notebook - u
{K EY1} (Kei)
05/25/2022, 1:49 AMah oki - u
{K EY1} (Kei)
05/25/2022, 1:49 AMit's basically set up the same way as the english notebooks i think - u
{K EY1} (Kei)
05/25/2022, 1:59 AMthis used japanese speech as input - u
{K EY1} (Kei)
05/25/2022, 1:59 AMthis used lines of syllables sung (for an utau bank) - u
{K EY1} (Kei)
05/25/2022, 1:59 AMthis used 100% computer generated singing syllables as input - u
{K EY1} (Kei)
05/25/2022, 2:00 AMand this used normal english as input. - u
user
05/25/2022, 2:50 AMOK good news and some bad news as well. I was able to get Coqui to train on CPU with no issues, however the training is very slow. I'm wondering if most of this has to do with the fact that I am not using a pre-trained model, but rather training completely from scratch. How much faster is using a pre-trained model versus doing it from the beginning? - u
{K EY1} (Kei)
05/25/2022, 12:47 PMIt's significantly faster tmk I've never done direct comparisons, though - u
user
05/25/2022, 2:03 PMThat's what I assumed. I just tried training Vits from scratch on the LJSpeech dataset, and got to only around 50 steps after 8 hours on CPU. There is a pre-trained model, but unfortunately no config.json to go with it, so coqui won't let me use it. - u
user
05/25/2022, 2:05 PMColab is an option, at least to train my own model that I can then fine-tune, but I'd rather do it all locally if I can. I need to get better hardware, that's for sure, but in the meantime I want to see if I can at least get somewhere with this. - u
{K EY1} (Kei)
05/25/2022, 2:57 PMah yeah, fair - m
mega b
05/25/2022, 3:10 PMCPU isn't the best idea if you are starting from scratch - h
hecko
05/25/2022, 3:13 PMstarting from scratch isn't the best idea in general - h
hecko
05/25/2022, 3:14 PMeven when training a different language it's better to have a starting point that can at least read left to right - h
hecko
05/25/2022, 3:15 PM,,hmm should i try training a fastpitch in coqui - h
hecko
05/25/2022, 3:17 PMor uh what's the funny diffusion one