I'm saying all of this assuming it connects via the main motherboard.

i mean phones already have chips specifically for ai or assuming it can make models usable on standard architectures (which isn't clear), it could be put in the cloud and rented out, like tpu·s

I don't think half of our training notebooks are designed to even allow TPU much less CPU only or whatever these chips are called

Neither do I think that the chips in AIs are designed to do anything except run AIs, not train them with massive datasets

the official article is very vague but i think it's talking about training https://blog.seas.upenn.edu/rethinking-the-computer-chip-in-the-age-of-ai/

> It supports on-chip storage, or the capacity to hold the enormous amounts of data required for deep learning, parallel search, a function that allows for accurate data filtering and analysis, and matrix multiplication acceleration, the core process of neural network computing.

cpu training is actually unintentionally supported, and i feel like tpu training should be easy enough but i've never done it so idk for certain

Is there a way to separate individual vocalists? I know about seperating vocals from instruments but I would really think an individual vocalist separator AI would be most useful..

How do you use CPU training? I'll do litterally anything no matter how long it takes to get around the stupid Colab limits. I've not figured out local setup on windows yet either

I really need a visual aid. I learn better visually.

if you wanna train on cpu on colab it's not worth it

it's like 10 times slower than gpu

and given that you get 3-4 hours of gpu per day it's slower than just doing that

but if you wanna know: runtime → change runtime type →

None

Most scripts in the talknet 2 training model don't work without a GPU. They'll complain about you needing (1) but I only have (0)

Most likely because it needs CUDA

Which I don't think you'll exacly get out of a CPU

oh talknet

i thought pipeline

What's the difference between pipeline and normal?

I came to the uberduck community after starting to learn talknet

pipeline is for tacotron 2

the main difference is that it supports multispeaker models

Oh alright

I'm just wondering I noticed my voices coming out of the uberduck website are noticably more compressed then they were when I was testing them. As in bitrate.

Is that true? Does uberduck compress?

not quite

the default output is 22.05khz but the synthesis notebook upscales that to 32khz so it's more like the notebook decompresses

So the notebook is automatically better? I find it is anyway because I can use the "reduce matallic noise (slow)" button which helps me loads. AND I can run the notebook for synthesis locally

Which, surprisingly, the local version of the notebook is much quicker than the online one.