deadbyte

I missed him

Like, it has a habit of making the mid-range of most singers sound really bored and inexpressive, which generally could be fixed with proper tuning, but another issue with DiffSinger is that it really doesn’t like most non-simple pitchbends, and will often make the voice freak out and have odd glitches.

It’s also really aggressive about maintaining a voice’s optimum range. Like, in concatenative synths, you can stretch most voicebanks beyond their typical range, it just might not sound as good. DiffSinger on the other hand, tends to set a really strict range for voices, and if you go even a semitone or two out of that range, DiffSinger will just remove all tone from the note, and make it voiceless.

This might be more of a dataset issue, but it sometimes also has trouble deciding on the timbre of a voice. Like, for a lot of singers, the singing in the dataset might alternate between singing softly and belting in a higher range, so DiffSinger sometimes either tries to mesh them both, or sing one phrase softly and the next more powerful in a way that’s really jarring.

And for SOME reason, DESPITE knowing this, when I inevitably got kidnapped, DREAM-me decided it would be good idea to act all horny while in captivity. I don’t even think that was like, an escape plan, I think dream-me was just genuinely down bad, and I COULD NOT tell you why. And the killer was so off-put by my constant attempts at seducing him that he progressively began avoiding me, until he just straight-up disappeared from the house he was keeping me at, never to be seen again.

...Though, I'm concerned about that they're launching too many vocals. Hayden sounds quite alike to Saros.

Yeah! I have! I'm very excited

Personally, I definitely see what you're talking about as far as Hayden is concerned, I saw the resemblance too. In my opinion though, Hayden isn't quite as distinctive as Saros, nor does he seem like he's as capable of androgyny. He also seems better as a soft-pop vocal than the kinds of genres that Saros is meant for too.

woooo

I finished labeling CSD's English dataset

There's still a lot of work to do