Wiki Wednesday #316 - Nagisa Maeno / 前野なぎさ

 it helps me to think of how much I have left to do in months instead of numbers!

Wiki Wednesday #316 - Nagisa Maeno / 前野なぎさ

Nagisa is an eighteen year old demigirl who uses she or they for pronouns. They are full of hope and optimism!

Who Is Maeno Nagisa?

Art from wiki

.

One of the people I work on stuff with hit me up and asked me to make them an English model from their Japanese data set. I was out of it and lazy... So I did it the "correct" way from before multidict training. I changed r to R and f to F in the English labels and wrote up a little replacements section for the English phonemizer.

It was great and all, but it kind of reinforced an idea that I just kind of rolled with and didn't bother challenging until I actually just did what was expected of me to begin with.

I prefer not using cross language synthesis in that way. When giving Japanese models English... I prefer to completely destroy the Japanese model's ability to use Japanese. I go through the labels and switch all of the vowels to Arpabet. I turn "r" into "dd". I go through and manually turn N into ng, n, and m. This sounds like effort and a bad idea. My vowel mappings are a, i, u, e, o to ae, iy, uw, ih, ah. As you may notice, if you're willing to say that i is not the same phoneme as i:, literally none of those are the same phoneme. That's kinda the point! Merging those phonemes allows the model to hang onto more of their accent. Wouldn't you expect a Japanese singer to have an unrounded u?

I did mention completely destroying the model's ability to use Japanese. Kinda, but also not. I haven't figured out how to make a dsdict that changes around the Japanese phonemizer. But as I've said before, I have beef with it. Using a Japanese model I've labeled and trained generally already requires using DIFFS as the phonemizer for the best results. I leave in all of the Japanese consonants and I have a dictionary file that allows the models to sing in Japanese as if it were still a Japanese model given you use DIFFS.

How to make your Diffsinger model sing in different languages is all up to you! I feel like I don't need to make a disclaimer that my way isn't the only way... I'm pretty sure a lot of people would outright say that my way is the wrong way given how well XLS (cross language synthesis) works when used as intended.

But in reality, I'm not very interested in cross language synthesis. I really just want a lot of English models! When it comes to music, I have really gravitated to only listening to English stuff. When I want to write a song, I want it to be in English. I love helping my Japanese friends make their own Japanese Diffsinger models, but that's mostly because it makes them happy. I have kinda no use for all of that! I just really, really like English.

How is Nagisa Maeno's bank?

Nagisa has a three pitch CVVC bank. Something felt really off. Her samples looked fine, but the renders were off. The pitch line in the oto editor was just... Missing in spots even though the sample was completely normal. 

I switched resamplers and deleted all of the FRQs. Just rerendering them using the default still broke them?! Their voice is lovely! I just don't understand what's going on.

Where Can I download Maeno Nagisa?

You can find them on their official site. This is a cool utau!