And who are those furry Lion and Deer guy behind them? Please tell me they are just back up dancers/Characters like that Pheonix and Cat human thing during the Aria Musical.
And who are those furry Lion and Deer guy behind them? Please tell me they are just back up dancers/Characters like that Pheonix and Cat human thing during the Aria Musical.
I'm not exactly a fan of furry stuff, but if they turned out to be vocal synths, I would still probably be happy about it because the lion one looks like he would have a deep, older male voice and we NEED more of those.
I legitimately don't think they're going to be singing synths, the guy talking during the stream said the girl was their 3rd ARIA (第三のARIA / daisan-me no ARIA) (this is what the Japanese users are callilng her right now) and was just like "Hmm, I wonder who these two guys are" playfully.
Oh yeah, I figured that they weren't important, I was just saying that I'd be perfectly indifferent to a vocal synth having a furry avatar if it had a decent male voice.
[IT media NEWS] made a cover testing Yukai, Kiritan and an unreleased Sasara in Cevio AI.
Anyone else kind of disappointed hearing that CeVIO AI has problems with Yukari? It makes me worried the other voice banks will have the same problems.
I read this blog by kM4osM (pronounced kurosu) where he explained common problems and (hopefully) how to fix them:
km4osm.com
Apparently long notes have pronunciation issues (singing turning muffled, becoming mispronounced, going off pitch, becoming quiet and loud again), which reminded me how we talked about KAFU having weird pronunciation/having to look at the lyrics, so i'm suspicious this will be a "CeVIO AI" problem. And I heard people saying that Yukari sounded best singing slow songs compared to fast songs, so uhh, sounds like Yukari working best with long notes will definitely make these problems crop up.
I've heard from 3 popular people in the fandom who own many synths from every engine that I trust say that Yukari goes off of pitch, which makes me worried because I do not have that good of an ear and having to manually fix a vocal synth which should be able to sing on better pitch than a human does not instill confidence. kM4osM said that if a note goes sharp you can literally move notes up/down to the neighboring note to compensate, change note length, try stuff like a "shi" note followed by "i" (same vowel), stick in breaths or slurs, try to put the funky notes on their own track, try using -, change location of consonants, use a っ, stick on ※ after a vowel to turn it into falsetto, add a bunch of the same vowel (ex: "aaaa" turns it into "a---"), pick a consonant + vowel that's sort of similar sounding instead, break apart contracted sounds (ex: "nya" into "ni + ya"), change TMG or ALP.
What do you all think? Personally, at the $200ish price tag, it seems troublesome to have those problems.
I read this blog by kM4osM (pronounced kurosu) where he explained common problems and (hopefully) how to fix them:
CeVIO AIの出音がなんか変だなーって思ったときに取る手段
CeVIO AI発売から約1週間ですね。みなさん麗さん使ってますか？このごろ、私のTwitterTLでは「麗さんはじゃじゃ馬」っていう話を聞くようになってきました。VOCALOIDや無印CeVIOを使ってた人の中にもCeVIO AIには手こ
km4osm.com
Apparently long notes have pronunciation issues (singing turning muffled, becoming mispronounced, going off pitch, becoming quiet and loud again), which reminded me how we talked about KAFU having weird pronunciation/having to look at the lyrics, so i'm suspicious this will be a "CeVIO AI" problem. And I heard people saying that Yukari sounded best singing slow songs compared to fast songs, so uhh, sounds like Yukari working best with long notes will definitely make these problems crop up.
I've heard from 3 popular people in the fandom who own many synths from every engine that I trust say that Yukari goes off of pitch, which makes me worried because I do not have that good of an ear and having to manually fix a vocal synth which should be able to sing on better pitch than a human does not instill confidence. kM4osM said that if a note goes sharp you can literally move notes up/down to the neighboring note to compensate, change note length, try stuff like a "shi" note followed by "i" (same vowel), stick in breaths or slurs, try to put the funky notes on their own track, try using -, change location of consonants, use a っ, stick on ※ after a vowel to turn it into falsetto, add a bunch of the same vowel (ex: "aaaa" turns it into "a---"), pick a consonant + vowel that's sort of similar sounding instead, break apart contracted sounds (ex: "nya" into "ni + ya"), change TMG or ALP.
What do you all think? Personally, at the $200ish price tag, it seems troublesome to have those problems.
I've been using her for a week and I can say with certainty that all these problems are very real and apparent, on top of these issues there seems to be dictionary issues? when singing do re mi fa so la ti do she sings do re me sa so la ti do? I think 200 for her is a bit... questionable. She's not versatile enough to justify it in my opinion.Anyone else kind of disappointed hearing that CeVIO AI has problems with Yukari? It makes me worried the other voice banks will have the same problems.
I read this blog by kM4osM (pronounced kurosu) where he explained common problems and (hopefully) how to fix them:
CeVIO AIの出音がなんか変だなーって思ったときに取る手段
CeVIO AI発売から約1週間ですね。みなさん麗さん使ってますか？このごろ、私のTwitterTLでは「麗さんはじゃじゃ馬」っていう話を聞くようになってきました。VOCALOIDや無印CeVIOを使ってた人の中にもCeVIO AIには手こ
Apparently long notes have pronunciation issues (singing turning muffled, becoming mispronounced, going off pitch, becoming quiet and loud again), which reminded me how we talked about KAFU having weird pronunciation/having to look at the lyrics, so i'm suspicious this will be a "CeVIO AI" problem. And I heard people saying that Yukari sounded best singing slow songs compared to fast songs, so uhh, sounds like Yukari working best with long notes will definitely make these problems crop up.
I've heard from 3 popular people in the fandom who own many synths from every engine that I trust say that Yukari goes off of pitch, which makes me worried because I do not have that good of an ear and having to manually fix a vocal synth which should be able to sing on better pitch than a human does not instill confidence. kM4osM said that if a note goes sharp you can literally move notes up/down to the neighboring note to compensate, change note length, try stuff like a "shi" note followed by "i" (same vowel), stick in breaths or slurs, try to put the funky notes on their own track, try using -, change location of consonants, use a っ, stick on ※ after a vowel to turn it into falsetto, add a bunch of the same vowel (ex: "aaaa" turns it into "a---"), pick a consonant + vowel that's sort of similar sounding instead, break apart contracted sounds (ex: "nya" into "ni + ya"), change TMG or ALP.
What do you all think? Personally, at the $200ish price tag, it seems troublesome to have those problems.
It really reminds me of that one earlier Kafu demo we've heard where we noted she sounds extremely pitchy.
Hopefully it's just a thing Cevio AI is prone to and only bad Cevio AI vbs take on these characteristics, Kafu had the excuse of still being in production at that time so I guess we'll have to find out.
Also when I and my girlfriend watched Yukari Rei tuning live stream ( just to see how the software works and how Yukari behaves because we were curious ) we did definitely note she was extremely wobbly and unstable-sounding.
The problem is that Techno-Speech themselves do the programming for the AI, you only need to provide them with hours of singing data and they do all the work themselves (even the phoneme labeling), so technically any "bad voice banks" would be mostly Techno-Speech's fault other than maybe blaming the voice provider for singing off key or something that was used for deep learning data. S:
(In KAFU's case, though, her having her voice changed to sound less like Kaf didn't help, but I'm kind of suspicious even if she sounded less high-pitched, she'd still have issues with muffled-ness and weird phoneme pronunciations but maybe not as noticeable.)
I initially looked at it as like "editor quirk" just like how Vocaloid tends to make voices more nasally and depending on a vb the nasally-ness is more or less severe but you're correct, it's just bad programming.
Uh oh, fingers crossed for IA AI. Lia has such a lovely voice it would be a real shame for it to be affected by an unstable engine.
I'm wondering if the quality issues are an AI, Techno Speech, or Cevio problem... Saki AI is still in her infancy, and she is more like a prototype than anything else atm. She doesn't have phoneme issues, but she does tend to have garbled and distorted vowels on certain high notes under very specific parameters. Yukari for Vocaloid is pretty high quality, but I found that the Vocaloid Engine and Editor make her sound very off pitch easily maybe due to how breathy she is, but being that the recording and resampling methods are different here, I wonder what this could mean.
From what I've seen of in editor usage, the amount of work arounds for glitches in Yukari Rei reminds me of the V2 era of using voicebanks like Lily and SF-A2 miki, its usable for the experienced voice synth user, but cumbersome and questionable as the technology is a decade newer and the price tag is hefty for a program that is seeing a more stable competitive program in SynthV roll out. Makes me concerned for Kiritan Cevio, and I can't even begin to think about c flower. v flower already has enough distortion going on in her vbs..
From what I've seen of in editor usage, the amount of work arounds for glitches in Yukari Rei reminds me of the V2 era of using voicebanks like Lily and SF-A2 miki, its usable for the experienced voice synth user, but cumbersome and questionable as the technology is a decade newer and the price tag is hefty for a program that is seeing a more stable competitive program in SynthV roll out. Makes me concerned for Kiritan Cevio, and I can't even begin to think about c flower. v flower already has enough distortion going on in her vbs..
If I can be honest... as Patuk mentioned, we watched a tuning stream a few nights ago. My issue isn't just with Yukari. It's the engine as a whole. Functionally, there's nothing new about it; it doesn't have anything CCS didn't have. The parameters don't respond any better than they do on the old engine. And the voicebanks, or at least Yukari, render warbley and off-pitch by default. And on a pettier note, it's outright hideous. You could at least make CCS look good, even though its default theme was an eyesore. It just looks cluttered and disorganized now.
Honestly? I'm baffled that it's being marketed as something new and exciting, and at such a high price tag. From what I've seen, I really feel it's a step down from CCS. I'm eager to see it improve and become great in the future, but right now, it's fallen waaaaaaaaaaaaaaay below my expectations, and then some.
Do keep in mind this is all just from the POV of an observer. I've yet to use it (because, well, I just can't find a reason to justify it right now), so take my opinion with a grain of salt, heh.
Honestly? I'm baffled that it's being marketed as something new and exciting, and at such a high price tag. From what I've seen, I really feel it's a step down from CCS. I'm eager to see it improve and become great in the future, but right now, it's fallen waaaaaaaaaaaaaaay below my expectations, and then some.
Do keep in mind this is all just from the POV of an observer. I've yet to use it (because, well, I just can't find a reason to justify it right now), so take my opinion with a grain of salt, heh.
CeVIO AI is a mixed bag, on the one hand it does a very good job at replicating a singer's voice - I don't think the timbre and nuances of Kafu's non-pitched voicebank could have been achieved with Vocaloid - but on the other this realism seems to create more chaos in the recordings.
AHS not supervising Yukari Rei's production seems to have really hurt her. All the other CeVIO AI Song voices don't seem to have the problems she has (as her major issues were apparent even in demos) so it makes me wonder what exactly went wrong. Part of me thinks it might be the fault of a lack of data but we can't be too sure without knowing exactly what they put in to her... I think she can probably be fixed but it's on VOCALOMARKETS to own up to it and initiate the process. I double checked and VOCALOMARKETS were definitely the ones responsible for producing the Song voice over Techno-Speech who seem to have had more of a supervisory role. As with SynthV I'd assume it's down to the companies to manage the labelling and then Techno-Speech handles the final run-over and packaging of the library.
Speaking of CeVIO though, KAFU's new demo actually sounds really good. Sounds like Police Piccadilly got to tune her too.
Speaking of CeVIO though, KAFU's new demo actually sounds really good. Sounds like Police Piccadilly got to tune her too.
I was gonna mention the new Police Piccadilly song here - very impressed with how Kafu sounds in that song, even though she sounds a bit muffled on the higher notes. (Though at this point, considering the issues surrounding Yukari Rei, I'm not sure if it's the VB or CeVIO AI to blame for the muffled sound.) I still also have that nagging feeling in my mind that the song could sound even better if Kafu was recorded at Kaf's natural pitch, but I'm starting to get over it now lol.
Techno-Speech required (if I remember correctly, don't want to dig for my old post with the time) 5 hours total of singing data. So I don't think it's a lack of singing data. If you compare it to NEUTRINO, I think that only had 1 hour's worth (at least for Kiritan when they released all that deep learning data for anyone to use)? And Synth V's Saki AI only was taught with 5 songs. So I feel like... maybe it's something Techno-Speech is doing differently than NEUTRINO and Synth V since the results are so drastically different?
Edit: That new demo with Police Piccadilly is the best demo I've heard with KAFU so far. Sounds cute and fun.
Seems like it's just VOCALOMARKETS messing up labelling then... Considering how the other CeVIO AI voices sound I don't think much blame can be put on Techno-Speech for Yukari considering she also has dictionary problems. With regards to Saki AI, I've been reliably informed the 5 songs is for her pitch model (the auto-tuning function) and she was trained on more than that for her actual database, though still much less than CeVIO AI. Just hoping they can do something to fix Yukari because I don't think anyone wants to pay $200 to use a standalone library like that.
