Kazuma Kiryu Singing AI test on NNSVS/ENUNU (Proof of Concept)

read pinned comment for more info
song by doriko
ust by haruvampire
special thanks to pix and canon, along with the yakuza modding discord server

[@Megurin-chan]

It legits sounds amazing-

[@CrowTenguOfBunbunmaru]

I want to see the Va of Kazuma sing this.

[@xFlormoon]

this is beautiful

[@haru0l]

*Please note that this is a raw output from the AI, the only thing I did was mix the cover and used the timing corrector. No pitch edits was made.* - - - This is a test of using low amounts of data (plus I wanted to hear how a jinriki ai vb would sound when made properly) due to the limited samples available after being ripped from the game. The model was trained using standard Conv1dResNet, *which is very prone to go off pitch* (I _do_ plan on training on ResF0 when I have the time which _should_ fix the off pitch and possibly allow for imitation of the singing style). The amount of data used in this is 12 minutes of audio (19 with silence) and the samples contains reverb which the AI hates (hence the noise on the consonants). The only other jinriki AI vb available would be HALC@LOIDs using samples of Haruka from the iDOLM@STER games which you can listen to here (kzbin.info/www/bejne/qKLYe42cdrqHq9E ). Also no I'm never releasing the model.

[@drawmuu]

I really appreciate you making the Google colab recipe for the training kit! I don't have high enough specs for it to work on my computer haha

[@huskytan]

love it

[@Purinkun]

This is so cool

[@ocuuda]

legit SO COOL just another source of inspo to actually get started on a nnsvs bank (+ i can only imagine how good he sounds with pitch editing and autotune oougugughhh)

[@haru0l]

hi. you dont have to imagine anymore

[@penixtrite6709]

HOLY SHIT

[@penixtrite6709]

You keep making the most amazing stuff and your tuning is excellent i wish more people could hear it aaaaaa

[@_Rhatsody]

Have no idea how ppl make these