# machine-learning
- s
Syclon (Creator of Quandale TTS)
05/03/2022, 4:42 PMAre TalkNet voices only available as referenced or non-referenced like Tacotron? - h
hecko
05/03/2022, 4:51 PMon uberduck only with reference audio on talknet they don't need reference audio but most of our models are barely trained for that so it might suck - u
{K EY1} (Kei)
05/03/2022, 9:38 PMDoes anyone else have trouble with tortoise tts where it sounds nothing like the reference clips I even fed it 215 wavs of the target voice and it just sounded like a mix between google assistant and ljspeech - c
Carson
05/03/2022, 9:51 PMyeah look at this i tried to do star butterfly and it turned into a generic girl voice - u
(Dawn) Will Draw Fictional Women
05/03/2022, 9:55 PMnot generic actor voice =/= tortoise no worky - u
{K EY1} (Kei)
05/03/2022, 9:57 PMOh. Really? Damn - u
{K EY1} (Kei)
05/03/2022, 9:57 PMI fed it lines from a literal speech corpus too and it didn't sound like it - u
{K EY1} (Kei)
05/03/2022, 9:58 PMThough i noticed that too with coqui Coqui replicates my voice, but not more quirky voices - s
Syclon (Creator of Quandale TTS)
05/03/2022, 10:00 PMWhat is the recommended amount of epochs that I should use when it comes to training in talknet, and not making the model sound like a robot? - g
Gosmokeless28
05/03/2022, 10:00 PMA transcription is not the same thing as a transcript, lol - u
{K EY1} (Kei)
05/03/2022, 10:00 PMIg you can pay attention to the val loss on that step where you change the epochs Make sure it hits 0.2 - u
{K EY1} (Kei)
05/03/2022, 10:01 PMI thought they were asking what to put for val filelist Which you put the path to the transcription txt - g
Gosmokeless28
05/03/2022, 10:11 PMYea, they were; I was just stating something unrelated - d
dzlandis
05/04/2022, 6:11 AMDid you mean to link the latest Talcotron instead of the latest TalkNet here since they were working with Talcotron? - d
dzlandis
05/04/2022, 6:12 AMJust wondering cause it's a little confusing 😅 - s
Syclon (Creator of Quandale TTS)
05/04/2022, 8:32 AMIn talknet, what makes the voice model better? The amount of epochs at step 9, or the amount of steps at step 11? - h
hecko
05/04/2022, 9:27 AMboth really the first makes the core of the voice sound better, the second polishes off some of the metallic tinge both have their limits though; for the first it's easy to tell because it'll eventually keep saying "not in top 3", the second doesn't have such an indicator but i think it's pointless to go above 5k steps unless you have a huge dataset - s
Syclon (Creator of Quandale TTS)
05/04/2022, 10:26 AMI have 93 wav files on the dataset. How many steps is it recommended to have? - h
haru0l
05/04/2022, 12:41 PMyou can try bruteforcing by increasing learning rate but it might sound bad - u
{K EY1} (Kei)
05/04/2022, 2:43 PMThe person who asked deleted their question. They were asking about talknet - u
{K EY1} (Kei)
05/04/2022, 2:43 PMAh, ok Thank you - u
{K EY1} (Kei)
05/04/2022, 2:43 PMI'll try that - s
Syclon (Creator of Quandale TTS)
05/04/2022, 3:24 PMI've heard that tortoisetts is kinda getting popular - s
Syclon (Creator of Quandale TTS)
05/04/2022, 3:25 PMWill it be featured as a model in uberduck? - m
matty1406
05/04/2022, 3:27 PMprobably not, unless it can be trained on better sources... tortoise is not very good with unique and non generic voices, and it is very slow to generate on as well - i
IBob012
05/04/2022, 3:28 PMtortoise is already pretty slow on colab, imagine it on uberduck with hundreds using it - m
matty1406
05/04/2022, 3:28 PMyeah, GPU takes around 5 minutes, but the colab warns if you use CPU to expect it to take up to an hour - u
{K EY1} (Kei)
05/04/2022, 6:05 PMI wish tortoise and coqui worked better with non-generic voices They have a really small scope of voices they work with, it's incredibly frustrating I just want easy crosslingual for japanese tts 😭 - m
myposynator
05/04/2022, 6:23 PMCould anyone tell me if this is correct? - u
(Dawn) Will Draw Fictional Women
05/04/2022, 6:37 PMthose graphs are actually pretty bad for a late epoch