Wiki Wednesday #375 - CoolJoule's Joule

 it is so tempting to ignore the old banks!

Wiki Wednesday #375 - CoolJoule's Joule

Joule is a cheerful fifteen year old who loves coffee. No gender is given on his wiki page, but he/him pronouns are used to describe him.

Who is Joule?

Art from wiki

.

I miss the Google Colab. I really liked training Diffsinger models in the Google Colab. Everything was easy, and most of all, it was aesthetic. That’s a fancy and incorrect way of saying that I like sans serif fonts and light mode.

But the colab was just broken. Not for anyone actually using Google’s cloud GPU. I saw multiple people successfully use it. But I just couldn’t get it to work with Docker anymore. I spent a few days sitting down and trying to make it work, but I seriously couldn’t.

So now I’m training locally locally and not just locally via Docker. And, as I said, I do not like it.

The colab makes it super easy and straightforward. It also means light mode and I can actually read how many steps I’m at. The command line? Maybe if I didn’t have such high magnification  giving me such big text I could see how many steps I was at, but I can’t see how many steps I’m at. I have to wait until it makes a checkpoint before I can see where I’m at… How do I know if I’m close to a checkpoint or not?!

And the handling of MultiDict… I am not a fan of it. I had to go into the config.yaml to say “I do not have a single speaker for Chinese or Japanese. This is an English model. Do not include Chinese or Japanese - there is literally no reason to include Chinese or Japanese!!”

But oh well. Whatever, I guess. If it works without paying anyone more money, that’s all I can ask for, right?

But yeah. Once I have it set up to run locally, there’s not much point in even opening Docker again. But I still want to. I really, really care about aesthetics when it doesn’t even matter.

Anyway, I actually opened up this computer to start working on more Diffsinger labels, but I was just like, hey, wouldn’t it be cool to get more work done for the blog? I need to work on that given I was reminded of it when someone told me FC2 was getting shut down.

Anyway, you may be able to tell that I’m manic. I’m all full of ideas and stuff but I get distracted and digress until I’m nowhere near where I need to be. I’m catching it. My excuse for leaving it in is just that like… I need words. I don’t actually need to make sure that each article has 600 words now that I don’t care about SEO, but I emotionally must make the effort because I gotta have some standards, right? At least I try to write good stories for Forgotten Friday. So there’s at least some weekly effort going on.

So yeah, I wish I could use the Google Colab to train Diffsinger banks but it’s broken and I don’t know how to fix it. (Other than, you know, actually using the cloud GPUs as intended.)

(Update: I now prefer difftrainer to the colab.)

How are Joule's banks?

Joule has four Japanese CVVC banks and one English CVVC Bank. Only the English Bank is considered current, but if I can touch a bank, I must use it!

There are two regular Japanese CVVC banks and two soft. 

The first CVVC bank is two mora with two pitches. He sounds very nice!

CVVC JA 2 has end consonants! There are five pitches. He is very expressive!

The prototype light Japanese CVVC has three pitches. He sounds nice and reassuring.

The Japanese alpha soft has five pitches, just like the CVVC 2. Until you realize... G4 is just an empty folder! He sounds great.

Finally, we have his only current bank: English CVVC beta. It only has one pitch. I... I actually have no X-SAMPA testers on my computer. Wait... This isn't X-SAMPA! Having an x-sampa ust wouldn't even help! I converted a C+V UST. He sounds great!

Where can I download Joule?

You can find him on his wiki page!