I've been missing from the forum lately, partly because I've been working on synSinger.

I had a prior version that worked fairly well, but it fell apart when it had to deal with fast tempo songs. I ended up rewriting most of the core code, and while it's still not completed, I can use it to render some simple examples.

Here's what it currently sounds like: Twinkle, Twinkle Little Star #11

It sounds better an octave lower: Twinkle, Twinkle Little Star #11a

For comparison, here's what synSinger sounded like before the rewrite: Twinkle, Twinkle Little Star #10

And finally, using Sinsy Twinkle, Twinkle Little Star - Sinsy

There's a clear improvement in Sinsy since I've last tried it, and clearly PG Music has put a bit of work into fixing the MusicXML output.

I'll be working next on getting synSinger glued back together so it can read MusicXML and work with BiaB again.

Excellent work David!
This must be incredibly challenging.
Keep it up.

Very nice, David. The 'accent' is less pronounced and the articulation is better in your product than Sinsy. I like the octave down as well. I'm not sure what that one is saying for "Twinkle", though.

I think the thing I dislike most about vocal synths is extreme legato from syllable to syllable. I wouldn't like it in an instrument sound, either.

Yes, there is lots to keep up with in Music XML, but if you get that working, your program would work with any music notation program that supports Music XML.

Nice work!

Yes, all these will be under user control. I'm sure uncontrolled under/overshoot would get old really fast, too.


Yep, there's a lot of tuning of the consonants that hasn't taken place yet. They're all quick and dirty version that need work.

I should also note that the only reason Sinsy speaks with an accent is because the training set it used had a person with an accent. I suspect that if it used a different training set, it would be pretty awesome. The Sinsy source code is available, but apparently is difficult to build, and only includes the Japanese data.


I've got a simple MusicXML parser written that reads the non-notation information, and that's sufficient for what I need. The real frustration was the MusicXML that BiaB had previously generated had a lot of errors - empty measures, durations not adding up correctly, incorrect accidentals on notes, rests inserted between syllables...

But it looks like it's working much better with Sinsy now, so I'm looking forward to having a much easier go at it this time around.

Thanks!

Well, humbling is more accurate. I'm just re-implementing existing technologies. Remember Software Automatic Mouth? And in the process of rewriting the code, I'm discovering a bunch of mistakes I made in prior versions.