grad-tts

I'm not super familiar with all the different models, I literally just installed it yesterday, but they seem to recommend starting with Tacotron DDC. I believe they have a pre-trained model for that, and when you download it it will explain where it's located on your computer so you can just point th fine-tuning arguments and give it your own data and configuration file. All of the stuff is in the documentation, and if you need links let me know.

i don't think coqui has grad-tts, that was a more general "what if i trained this architecture"

thankfully the grad-tts repo has a file called

train_multi_speaker.py

so i don't think it'll be too hard

Ah I see. Will look up Grad TTS. What are some of its advantages over something like TT or Vits?

it uses diffusion and diffusion is magic

in terms of listable benefits uhh you can control the speed, you can control the quality by diffusing less or more, control the intensity of intonation with temperature,

Woh. Will definitely check this out. Found instructions on how to finetune a model, so I'm going to attempt this and see how much of a speed increase I get.

Oh that's cool...

Ok trying to finetune a multi speaker vits model on LJ. No idea if this will work, but it looks like it's training.

Vits my beloved

I have no idea, I feel like I'm doing this wrong LMAO. Not getting any steps yet, and I've let it sit for like 10 minutes. Seems like the training process started though. Computer was faning like crazy earlier.

maybe should try glow TTS with fine-tuning instead? Don't know. I'm gonna let it sit for a little longer.

Vits is super resource intensive Like, on a good gpu it can take weeks

Ok, trying to train coqui on my own data, but keep getting list index out of range. Was able to do this with TalkNet so not sure what is happening. Worked fine on LJSpeech and this data is formated the same way.

it might be expecting a speaker id or some other extra data alternatively you might have a blank line

> Each sample is a list of

[text, audio_file_path, speaker_name]

I thought the text came after the audio file name. That's how the LJ speech dataset looks when I am reviewing it, and I have my data set formated the exact same way. Not really sure what's going on here, I might take a break babe for now and go ask on the Coqui github.

is that like the speaker_id thing for multi-speaker models? like

wav.wav.|Lorem ipsum|LjSpeech

instead of the 0?

ye im gonna have to look Coqui up

Anyone else having trouble with tensorboard not reading logs recently?

Ok think I got it. You need to have the dataset look exactly like LJ, or it won't work. This means having two versions of your transcript, although not quite sure what the differences between the two are just yet. Will try and train again maybe tomorrow. Oh BTW decided to make a new discord cause not a big fan of the username lmao, so will temperarily bee leaving but will come back soon.

Aren't you able to change your discord name under settings?

Should I be concerned if the curve sort of looks like this?

no

just means a shorter wav

heard there's a new version of ai were you can use your favorite artist and its kinda like the dragon program were you would sing the lyrics in your voice WITH your emotions anger etc BUT it would sound like the artist voice instead. I know it was in development forgot what it was called but I wanna make an athentic biggie album with featured artists such as 2pac and others including myself.

working on making a dataset for Buffy Summers. i've been tampering with it for the past couple days in audo/uvr5. is this clean enough to use or should I scrap anything with even remotely audible bg noises

Talknet can do that