Suppose you come up with lyrics, cover art, and other original ideas... You still have yet to get this to the finished stage where at least a master copy of a CD exists. You know, do the singing, then all the studio work... I'm no singer, but making print-worthy cover art's not that easy either - I've done it, and puzzling through the printer's requirements took a while by itself, satisfying them took even longer.
They're planning to convince kids that all it takes to make an album is to make lyrics, cover art, and poof, here it is, somewhere? That's plain sick.
HP Jornada 720 handheld PC 2000 with Windows CE comes with a copy of Yahoo! Messenger in rom. It doesn't work now - Yahoo did something to the server so it stopped logging in about a year ago. The machine's not three years old yet and is still one of the best clamshell handhelds out on the market, and here it's carrying dead weight in ROM, at no fault of HP. Unlike more modern PocketPCs, it doesn't have flash rom so it can't be overwritten.
It can't be replaced at the cost of normal memory either, because HPC (or even PocketPC) versions of Yahoo! Messenger are not available anymore, they simply vanished about the same time when it stopped working, without as much as a mention or explanation.
Needless to say this didn't improve my opinion of Yahoo. The ONLY free IM client for HPC 2000 that still works and properly handles cyrillic is AIM. And the American version is not free, British is. (!)
P.S. And the only Jabber client for HPC is both non-free and really really sucks.
They are using what they call ARPAbet, which is a phonetic alphabet originally developed by DARPA and seems to be the result of that Navy work you mention. I suppose it was used in some original theoretical work, maybe in one of the papers by Klatt on who's work all the formant synthesis systems seem to be based, but I can't get at his papers. (and if I could, the problem of making the missing link in the TTS system, the phoneme->Klatt synth parameters translator, would be relatively easy to solve)
Are you sure? Where does this information come from? SoftVoice's site says nothing about Navy, but does mention MacinTalk and the Amiga narrator.device being the same thing in different versions.
I've been to all three a while ago, actually, but thanks anyway.:) Here are the results of my findings on those:
'say' is also known as rsynth. I tried getting it to work. It works. I even made it compile and work (for suitably small quantities of 'work') under DJGPP+Allegro. But it is coded in a very messy manner, with sections of the code commented out for no listed reason. It is not clear how it works, using it to build upon is... well, it is for someone who knows the phonetic theory behind them better than I do, at the very least.:) It compiles, it runs, but it's nowhere near what Amiga did. Barely legible, even. Improving it turns out much harder than it looks.
The Klatt 3.04 package looks very promising, but it is not a complete text-to-speech system. It's just the last part of the tract, where it converts 'vocal tract parameters' into sound. The question of how to get vocal tract parameter sets from known phonetic representation of a phrase in ARPAbet is not solved by this package. It'd be nice if I could just stick it on the end of the rsynth, but the parameter sets don't match and I don't know which are which!
As for the implementation that's a plugin for the 'pure data' package, I can't even make heads or tails of what the goddarn package is supposed to do - let alone how to use the library planned for it to make a speech driver...
You probably don't need a complete, general-purpose voice for home automation.
That's only if we're talking about dumb home automation. What if I want it to read email headers aloud to me? Give me names of people calling my phone? Amiga did that just fine.
Amiga's voice driver was based on formant synthesis, not diphone synthesis. As a result, it might have been worse in sound quality, (it was, true) but diphone synthesis can't compare to it in flexibility. It didn't need complex voice definitions, a voice definition was just a set of ten or so Klatt synth parameters. I also don't know any other synth that would have such an easy time speaking English, Russian and Japanese in a single sentence. And fit in under 64k compiled code...
That, by the way, is precisely the best explanation for why we hear the laser blasts in space opera movies - because onboard computers synthesise engine and blast noises to help the pilot, even though there's no sounds in vacuum.
I've seen Ananova. I don't have enough money for the L&H engine.:) Which is what's so annoying about the whole thing - we can't do something I had working perfectly well for no extra penny than what was spent on the machine itself, on an ancient Amiga 500 with measly CPU speeds and no hard disk with our modern machines that outrun and outsmart the poor bitty box many times over. I even thought about emulating it and leaving it like that, but I can't even read the Amiga floppies in PC drives.
So far, my best hopes rest with using the various free implementations of Klatt synthesisers with existing code snippets that do the translation from text to phoneme sets. Unfortunately, the piece in the middle, the code that converts phonemes into Klatt synthesiser input data, is missing and I can't find anything on the net - not even the bit that they used in making rsynth and copied from some book, mangling it throughly in the process.
I thought about MBROLA. There's nowhere in the pipeline of a diphonic synth where I could stick a check for which phoneme is currently being sounded. Therefore I can't match phonemes to visemes and can't do lip synch.:( Besides last time I checked, MBROLA had only two female voices...
Actually, lip sync was a feature of the speech driver on Amiga! At any moment you could get two variables for mouth height and width connected to the currently sounding phoneme, so all it took was several sprites depicting some characteristic lip positions and an eternal loop. No idea where the synth driver got those from, but my current theory is that it just had an array of viseme data corresponding to phonemes and would point to it while passing the phonemes to the final stage in the Klatt synth. Programming the voice just took an array of about ten variables like on other Klatt-based parametric speech synthesisers. Most of the hundred lines dealt with sensors, since it was so convenient to use Amiga joystick port to grab pin states.
When I come back home and my computer greets me, cause there's nobody else to do that, it fels much better when it's a female voice, pod bay door or no pod bay door.:)
As far as I know, the synth used in MacOS is the direct descendant of the synth used in AmigaOS up until they lost the license, the next generation of SoftVoice family.
What irks me about Linux home automation is the absence of a good speech synthesiser. I mean, I can automate everything, stick sensors everywhere, make control agents, a microcontroller network, but the endearing thing about it was that my computer/talked/ to me.
Festival doesn't quite cut it, not because sound quality is bad (it is quite good) but because making new, custom voices for it is a laborous, time consuming process and no good female voices are available. And doing lip synch with it is something I don't have a fscking idea about how to do. On my old Amiga 500 that was just a matter of drawing some sprites and a hundred of lines of AMOS Basic, no more.
I switched to Linux for my home server from OS/2 specifically because I wanted to replace my dying Amiga 500 that was the voice agent for my home automation, and figured that I'd have better chance to find good free speech synthesis for a more modern, free OS. That was, like, almost two years ago.
Amiga finally died but I/still/ can't pull anything like this off. What I could assemble from public sources is painfully crude even when compared to the ancient Amiga SoftVoice/Nowspeak.
Don't even remind me about rsynth...
Slashdot Mirror
Suppose you come up with lyrics, cover art, and other original ideas... You still have yet to get this to the finished stage where at least a master copy of a CD exists. You know, do the singing, then all the studio work... I'm no singer, but making print-worthy cover art's not that easy either - I've done it, and puzzling through the printer's requirements took a while by itself, satisfying them took even longer.
They're planning to convince kids that all it takes to make an album is to make lyrics, cover art, and poof, here it is, somewhere? That's plain sick.
P.S. No, I can't RTFA. Damn NYT...
That's it, I'm swearing off speaking this one and switching to Esperanto. At least the amount of people eligible to copyright it is much smaller. :)
HP Jornada 720 handheld PC 2000 with Windows CE comes with a copy of Yahoo! Messenger in rom. It doesn't work now - Yahoo did something to the server so it stopped logging in about a year ago. The machine's not three years old yet and is still one of the best clamshell handhelds out on the market, and here it's carrying dead weight in ROM, at no fault of HP. Unlike more modern PocketPCs, it doesn't have flash rom so it can't be overwritten.
It can't be replaced at the cost of normal memory either, because HPC (or even PocketPC) versions of Yahoo! Messenger are not available anymore, they simply vanished about the same time when it stopped working, without as much as a mention or explanation.
Needless to say this didn't improve my opinion of Yahoo. The ONLY free IM client for HPC 2000 that still works and properly handles cyrillic is AIM. And the American version is not free, British is. (!)
P.S. And the only Jabber client for HPC is both non-free and really really sucks.
They are using what they call ARPAbet, which is a phonetic alphabet originally developed by DARPA and seems to be the result of that Navy work you mention. I suppose it was used in some original theoretical work, maybe in one of the papers by Klatt on who's work all the formant synthesis systems seem to be based, but I can't get at his papers. (and if I could, the problem of making the missing link in the TTS system, the phoneme->Klatt synth parameters translator, would be relatively easy to solve)
Are you sure? Where does this information come from? SoftVoice's site says nothing about Navy, but does mention MacinTalk and the Amiga narrator.device being the same thing in different versions.
I've been to all three a while ago, actually, but thanks anyway. :) Here are the results of my findings on those:
:) It compiles, it runs, but it's nowhere near what Amiga did. Barely legible, even. Improving it turns out much harder than it looks.
'say' is also known as rsynth. I tried getting it to work. It works. I even made it compile and work (for suitably small quantities of 'work') under DJGPP+Allegro. But it is coded in a very messy manner, with sections of the code commented out for no listed reason. It is not clear how it works, using it to build upon is... well, it is for someone who knows the phonetic theory behind them better than I do, at the very least.
The Klatt 3.04 package looks very promising, but it is not a complete text-to-speech system. It's just the last part of the tract, where it converts 'vocal tract parameters' into sound. The question of how to get vocal tract parameter sets from known phonetic representation of a phrase in ARPAbet is not solved by this package. It'd be nice if I could just stick it on the end of the rsynth, but the parameter sets don't match and I don't know which are which!
As for the implementation that's a plugin for the 'pure data' package, I can't even make heads or tails of what the goddarn package is supposed to do - let alone how to use the library planned for it to make a speech driver...
You probably don't need a complete, general-purpose voice for home automation.
That's only if we're talking about dumb home automation. What if I want it to read email headers aloud to me? Give me names of people calling my phone? Amiga did that just fine.
Amiga's voice driver was based on formant synthesis, not diphone synthesis. As a result, it might have been worse in sound quality, (it was, true) but diphone synthesis can't compare to it in flexibility. It didn't need complex voice definitions, a voice definition was just a set of ten or so Klatt synth parameters. I also don't know any other synth that would have such an easy time speaking English, Russian and Japanese in a single sentence. And fit in under 64k compiled code...
That, by the way, is precisely the best explanation for why we hear the laser blasts in space opera movies - because onboard computers synthesise engine and blast noises to help the pilot, even though there's no sounds in vacuum.
I've seen Ananova. I don't have enough money for the L&H engine. :) Which is what's so annoying about the whole thing - we can't do something I had working perfectly well for no extra penny than what was spent on the machine itself, on an ancient Amiga 500 with measly CPU speeds and no hard disk with our modern machines that outrun and outsmart the poor bitty box many times over. I even thought about emulating it and leaving it like that, but I can't even read the Amiga floppies in PC drives.
So far, my best hopes rest with using the various free implementations of Klatt synthesisers with existing code snippets that do the translation from text to phoneme sets. Unfortunately, the piece in the middle, the code that converts phonemes into Klatt synthesiser input data, is missing and I can't find anything on the net - not even the bit that they used in making rsynth and copied from some book, mangling it throughly in the process.
I thought about MBROLA. There's nowhere in the pipeline of a diphonic synth where I could stick a check for which phoneme is currently being sounded. Therefore I can't match phonemes to visemes and can't do lip synch. :( Besides last time I checked, MBROLA had only two female voices...
Actually, lip sync was a feature of the speech driver on Amiga! At any moment you could get two variables for mouth height and width connected to the currently sounding phoneme, so all it took was several sprites depicting some characteristic lip positions and an eternal loop. No idea where the synth driver got those from, but my current theory is that it just had an array of viseme data corresponding to phonemes and would point to it while passing the phonemes to the final stage in the Klatt synth. Programming the voice just took an array of about ten variables like on other Klatt-based parametric speech synthesisers. Most of the hundred lines dealt with sensors, since it was so convenient to use Amiga joystick port to grab pin states.
When I come back home and my computer greets me, cause there's nobody else to do that, it fels much better when it's a female voice, pod bay door or no pod bay door. :)
As far as I know, the synth used in MacOS is the direct descendant of the synth used in AmigaOS up until they lost the license, the next generation of SoftVoice family.
What irks me about Linux home automation is the absence of a good speech synthesiser. I mean, I can automate everything, stick sensors everywhere, make control agents, a microcontroller network, but the endearing thing about it was that my computer /talked/ to me.
Festival doesn't quite cut it, not because sound quality is bad (it is quite good) but because making new, custom voices for it is a laborous, time consuming process and no good female voices are available. And doing lip synch with it is something I don't have a fscking idea about how to do. On my old Amiga 500 that was just a matter of drawing some sprites and a hundred of lines of AMOS Basic, no more.
I switched to Linux for my home server from OS/2 specifically because I wanted to replace my dying Amiga 500 that was the voice agent for my home automation, and figured that I'd have better chance to find good free speech synthesis for a more modern, free OS. That was, like, almost two years ago.
Amiga finally died but I /still/ can't pull anything like this off. What I could assemble from public sources is painfully crude even when compared to the ancient Amiga SoftVoice/Nowspeak.
Don't even remind me about rsynth...