Uploaded September 2025 | Updated September 2026, 1 day ago

---

Vocals and lyrics are AI generated. All MIDI notes were hand placed by me and all audio other than the vocals was rendered using VSTis that don't use AI in REAPER.

I love my silly little midi keyboard. Using REAPER on my Windows tablet is really annoying, but being able to input notes via step mode by pressing piano keys is amazing!

This is an Utau to Diffsinger conversion using my own Nicky B. Cantonese bank. I cannot speak Cantonese. However, this differs from a normal conversion because the variance model was trained on text to speech output. It's a weird legal grey area because no TTS audio was used in training - only analysis of it.

Anyway, the method was fun to try but I won't do it again. Even though I was using one of the most advanced Cantonese TTS things out there, it still ended up being robotic.

and on the Utau front, it's just difficult. This model cannot sing long notes because long notes were not in her training data - that's why all of the notes here are so short. (okay, and the fact that step mode means you need to change the grid to change the length of the note...)

Seriously, the song was meant to have a BPM of 110 and I said "whoops!" and changed it to 140!

so yeah!! I'm really happy I did this experiment. If I ever try Cantonese again, I already have a great dsdict for it!