Yeh - not sure this will help
however - a (very talented) colleague of mine worked on a simple speech
recognition software for mobile - it was built to recognise about 20 commands
with 90% success rate.
His approach (in my simplistic terms) was:
1) get recordings / audio samples of the
I've started implementing some code this morning in the hopes to match the
vowel a this morning. Of course there are several intonations for this
depending on the word it's located in, but if I can get a match on a naked
a I may be on to something. Like you said, I have a higher chance of
success
I can get waveforms... but say a takes 1 second to speak. I get different
waveforms over that 1 second... so I'm not matching against a single
waveform, but many waveforms in succession. This seems like a tricky thing
to match against.
What might be a good approach to matching values over a
I was able to match a single a - although even with a straight a there
can be some subtle variation. So I mapped variations that come close and I
don't need to match every value in the complete waveform over time... every
couple together or even the first value with buffer comes pretty close.
I would say there are about 5 - 7 mouth shapes you could distribute
through your animation that would give the impression that the avatar
is saying the right words.
Plus if your animation is fluid (meaning it doesn't look like the
avatar is straining to say the words) it probably wont be
If your mp3's are pre-recorded rather than people recording them
dynamically, could you use cue points?
On 02/06/2010 20:57, Eric E. Dolecki wrote:
I have a face that uses computeSpectrum in order to sync a mouth with
dynamic vocal-only MP3s... it works, but works much like a robot mouth. The
You could try matching say a lowered jaw with low octaves and a
cheeky jaw with high octaves.
JAT
Karl
On Jun 2, 2010, at 3:20 PM, Eric E. Dolecki wrote:
This is a software voice, so nailing down vowels should be easier.
However
you mention matching recordings with the live data. What is
It's using dynamic text to speech, so I wouldn't be able to use cue points
reliably.
On Thu, Jun 3, 2010 at 4:09 AM, Glen Pike g...@engineeredarts.co.uk wrote:
If your mp3's are pre-recorded rather than people recording them
dynamically, could you use cue points?
On 02/06/2010 20:57, Eric
I don't think that's enough. Has anyone seen pitch detection in AS3 yet (no
microphone source)? That might be enough but I'm not sure.
On Thu, Jun 3, 2010 at 5:55 AM, Karl DeSaulniers k...@designdrumm.comwrote:
You could try matching say a lowered jaw with low octaves and a cheeky jaw
with
Eric E. Dolecki wrote:
It's using dynamic text to speech, so I wouldn't be able to use cue points
reliably.
Use dynamic cuepoints and stop complaining. If it can generate voice, it
can tell you what kinds of voice it put where. It is far more exact than
trying to reverse the incredibly
I've tried running software voice vowels through the system and I am able to
create signatures for the vowels that's somewhat accurate (depending on
how it's influenced in a word or if it's standalone). I've run them several
times and my values always seem to match (which is good). I end up with a
My most humble apologies go out to Henrik and anyone else who felt that I
was complaining about something. Which I wasn't.
On Thu, Jun 3, 2010 at 10:02 AM, Henrik Andersson he...@henke37.cjb.netwrote:
Eric E. Dolecki wrote:
It's using dynamic text to speech, so I wouldn't be able to use cue
My most humble apologies go out to Henrik and anyone else who felt that I was
complaining about something. Which I wasn't.
I don't think you need to apologize, I didn't think you were complaining at all
- just stating your view of how you see this technique working with your
project.
Jason
Have a look at this stuff, don't know if there is source, but it might
be a start...
http://www.allflashwebsite.com/article/real-time-lip-sync-in-flash
On 03/06/2010 15:03, Eric E. Dolecki wrote:
I've tried running software voice vowels through the system and I am able to
create signatures
Wow. That was really uncalled for.
Anyway, if you can pre-generate samples for all vowels for all samples, I
can't see why comparing them to the speech generated by the same system
would be any harder than comparing it to a number of collected profiles.
You really just need to collect profiles
I'm abandoning the whole vowel recognition unless I can find something
someone else has done to base my implementation on. I've burnt too much time
on it for something that won't give a whole lot of bang for the buck. It's a
very complex problem (for me anyway).
Eric
On Thu, Jun 3, 2010 at 2:37
Juan Pablo Califano wrote:
Wow. That was really uncalled for.
I meant no ill will. I just meant that there is a better solution than
this idea.
For me, the hard part, which you seem to imply is rather simple here, is
*matching+ the input audio against said profiles. Admitedly, I don't
Juan Pablo Califano wrote:
Wow. That was really uncalled for.
That was my reaction, too. I didn't see Eric as complaining--just
asking. Maybe Henrik was just having a bad day.
For me, the hard part, which you seem to imply is rather simple here, is
*matching+ the input audio against said
I meant no ill will.
I just meant that there is a better solution than this idea.
Ah, so that's what you meant by stop complaining. I see, that makes
perfect sense now. ;)
Jason Merrill
Instructional Technology Architect
Bank of America Global Learning
Join the Bank of America Flash
there was a really good commercial lip-synch program that generated
Action Script to control mouth positions.
I don't know if it's still around--that was 5 years ago,
and it was pretty expensive (about $2,500 for one seat, I think).
It may even have
been a Director Xtra that worked with a
I'm not Henrik, but I've done some lip-synch stuff for Disney. We did
it pretty much the way Eric described--we just used amplitude. It's
not as accurate as Disney would demand on a film, but it's ok in the
kids' game market.
I see, amplitudes could be just good enough for some stuff.
Ya - I have the data for both things, but they extend over time and are
difficult to compare. It's the boiling down the signatures into something
simple and being able to read the playing audio looking for the match (or
near match). I thought about using bitmap data and trying to match up
Jason Merrill wrote:
You're probably thinking of the Flash-based SitePal:
http://www.sitepal.com/ ?
It could have been. I honestly don't remember--it was at least 5 years
ago. We considered using the software, but the studio head vetoed it
as too expensive, especially since we already had an
Don't know if this will help, but have you looked into
WaveAnalyzer.as or
Flash MX - Audio: Sound completion event (The source files for this
can be found in the Flash MX/Samples folder.)
They both let you control the sound. I am thinking this will point
you in a good direction. Its AS2
If you need any of these files or can't find them, lmk and I can send
off list.
Best,
Karl
On Jun 3, 2010, at 3:37 PM, Karl DeSaulniers wrote:
Don't know if this will help, but have you looked into
WaveAnalyzer.as or
Flash MX - Audio: Sound completion event (The source files for this
I think I might make waveform bitmaps and then try and compare against the
current waveform (block EQ) - and if it's a close match, then fire off
specific vowel events. If that works, I could do consonants too. If this
works, I'll do jumping jacks and shots of Jack.
So how would I compare two
Do you have SoundEdit? Or the like?
Karl
On Jun 3, 2010, at 5:09 PM, Eric E. Dolecki wrote:
I think I might make waveform bitmaps and then try and compare
against the
current waveform (block EQ) - and if it's a close match, then fire off
specific vowel events. If that works, I could do
Before you start reinvesting the squarewheel, at least do some research
on how people are doing it.
I did not learn enough from it personally, but I can tell that it is a
good book:
http://www.dspguide.com/pdfbook.htm
Read it and then do the matching algorithm. This way you will avoid
SoundBooth
On Thu, Jun 3, 2010 at 6:39 PM, Karl DeSaulniers k...@designdrumm.comwrote:
Do you have SoundEdit? Or the like?
Karl
On Jun 3, 2010, at 5:09 PM, Eric E. Dolecki wrote:
I think I might make waveform bitmaps and then try and compare against the
current waveform (block EQ) -
Dude, whether you know it or not, you come off being pretty arrogant with
your comments. Don't worry, I won't be ending up on the dailywtf anytime
soon.
On Thu, Jun 3, 2010 at 6:36 PM, Henrik Andersson he...@henke37.cjb.netwrote:
Before you start reinvesting the squarewheel, at least do some
I would try using that to figure out a way of maping the sounds and
then translate that to your project. You are able to see the wave
forms in soundbooth? Haven't used it. If so, can you run your cursor
over it at any point to get the readings? Might be a little trivial,
but may yeild a
Eric E. Dolecki wrote:
I have a face that uses computeSpectrum in order to sync a mouth with
dynamic vocal-only MP3s... it works, but works much like a robot mouth. The
jaw animates by certain amounts based on volume.
I am trying to somehow get vowel approximations so that I can fire off some
This is a software voice, so nailing down vowels should be easier. However
you mention matching recordings with the live data. What is being matched?
Some kind of pattern I suppose. What form would the pattern take? How long
of a sample should be checked continuously, etc.?
It's a big topic. I
33 matches
Mail list logo
