TF2

listen to it

listen to the tone, the vocal range, the delivery

Umm this showed up after putting the wavs.zip in the auto transcriber. What do I do?

Oh wait it's still going.

there's an exclusive contributors-only notebook called

link2model

that does exactly that given a youtube video it sucks though so you're not missing out on much

Instant pot sushi rice - the AI is hungry lol

where's the link to download everything that would required and the github repo or videos guide

I don't know if this is the right place to post this, but the Talknet Training notebook has stopped working. I've tracked the error and it's due to an update in the 'omegaconf' package that happened a couple of days ago. Easy fix by just using the previous version of that package.

Glad I'm not the only one who noticed this. How would I go about getting the previous version of that package?

Not exactly machine learning but I do like that somebody actually did this https://colab.research.google.com/github/o-date/photogrammetry/blob/master/Meshroom_%2B_GPU_for_Photogrammetry.ipynb#scrollTo=wtz12QNl930T

^ a colab for meshroom. GOod if you don't have a GPU to run it on.

OK I know I'm really late to the party, but I just tried the Uberduck Pipeline training notebook for Tacotron with the ability to listen to the model as it trains, and it's a game changer. After I figured it out, it was huge. I'm totally blind so I can't see the graph, but listening to the audio helps me to figure out when I should stop training. Would it be at all possible to add this to the TalkNet notebook as well, as I find that model to be easier to train along with needing less data. Thank you guys so much for democratizing custom text to speech voices in this way, I'm so excited to dig into this further.

the talknet notebook was not made by us so i'm not sure how practical it would be to modify it when i train it i usually check how many of the last few lines contain the words "not in top 3", meaning it hasn't improved during a given epoch but it might be quite awkward to do with a screen reader

i do commend you for even attempting to train a model as a blind person, let alone succeeding

Thanks man, Colab is pretty accessible seeing as it's made by Google, so looking at the status isn't really a problem, I just like having the peace of mind. I've only begun training TalkNet relatively recently, but have gotten much better results than with Tacotron so figured I would ask. I'm currently attempting to train a Tacotron model but it's taking significantly longer than TalkNet. Unfortunately seems like the TalkNet notebook is broken currently, so in the meantime I thought I would give Tacotron another try after not having used it for like a year.

Yeah, I thought talknet notebook broke again. I had my dataset ready for this.

A little back in the message history someone gave a solution to fix this, but I'm not 100% sure how to implement it as I'm still kind of a Python noob lol.

Ah ok lol

Epoch 10 from a retrain of my Imogen Model with a slightly trimmed list.

Think i'm going to have 2-4 seperate Imogen datasets if I feel like it. a "full" one of like up to 500 files per wav folder, normal speaking and singing, then a "sample" or wahtever of half the larger versions. The new one in training("cut_382.wav") is definetly an improvement over the one on uberduck(uber.wav")

Hey man, first of all love Imogen Heap, and the model sounds great. Are you using TalkNet for this?

it's a Tacotron2 model. Never can get Talknet to work

Just tried TalkNet a few days ago and one of the notebooks I found didn't work, but I found another one that updated some Python packages and it works pretty well. I believe that's under the notebooks category, and it's a back up of the TalkNet training notebook. I have the opposite problem, I could never get my Tacotron models to produce intelligible speech. I just discovered the pipeline training notebook with the ability to listen to the model as it trains, so hopefully tomorrow I'll give it another shot and see if I can get somewhere. I believe currently the TalkNet notebook might be broken, you need to downgrade a Python package or something, but I have not looked into it yet.

When (or if, knowing me) I get around to sorting the dataset into the seperate types, it maybe good to try Imogen with Talknet. At any rate, maybe somebody that knwos how to run it right can ru nit.

Feel free to send the dataset to me and I'd be more than happy to train it for you. I'm still pretty new to all of this, but more than happy to help out in any way I can.

Training jus quit ~1.2 loss for I think 366 so far. Sounds okay for what I have so I'll send it

(or upload to uberduck. whatever)

???

Kinda sounds like cookie monster lmao.