Amazon: IVONA bevy of 'all natural' blabber babes to beat Siri Amazon bought text-to-speech company IVONA systems on Wednesday, the online book-floggers have announced. The acquisition fuelled rumours that Amazon, the quietest member of The Gang of Four, is planning a rival to Apple's talking assistant Siri. Amazon announced the deal with Ivona Software yesterday on its website, but didn' …
This is a fun article, about how Siri grew out of a DARPA project, was intended to do much more than she does now, and after Apple bought her (snatching her away from Verizon's Android handsets) they curbed her abilities and potty mouth. She seems to have plenty of cousins, though.
http://www.huffingtonpost.com/2013/01/22/siri-do-engine-apple-iphone_n_2499165.html
Over a short span, one of the U.S. female voices is quite indistinguishable from real (although one of the male voices does not quite convince) in that if I did not know beforehand, I'd assume it was real.
However, natural reading of various TYPES of texts (news, non-fiction, novels) requires adjustment. The voice for a Jane Eyre or Mickey Spillane reading would really have to change from that for a science book. Even a dull Librivox speaker will put some life into the characters. You just can't do that with TTS, probably never will. One has to understand each person in the story, and the inflection and modulation also changes from sentence to sentence. Affect is something inferred from the text, it cannot be calculated by an algorithm.
Even a dull Librivox speaker will put some life into the characters. You just can't do that with TTS, probably never will.
"Never will" is rather too strong. While we're still very far away from a decent understanding of affect, much less a formal model of it that would let us implement it algorithmically, a lot could be done in this area with predictive models. NLP researchers have made significant progress in recent years in systems that can model rudiments of narrative; going from there to building decent predictive models of the purported emotional states of characters is not a great leap. They wouldn't be nuanced, and no one would claim this represents "understanding" emotion, but if the model predicts a valid emotional overtone for the subject text most of the time, that's good enough for TTS purposes.
Then it's a matter of varying the TTS prosodic parameters to convey that emotional overtone. That's tricky in itself, because prosody is not terribly well understood itself; for example, linguists who study prosody still can't agree on what in English pronunciation actually conveys "stress". But again you can build a pretty effective predictive model without fully understanding the domain being modeled.
Affect is something inferred from the text, it cannot be calculated by an algorithm.
That's a fallacious argument, unless you can prove that human readers don't "calculate" affect "by an algorithm".
I tried a couple of voices on: "I can say fairly complex, as opposed to complicated, sentences without stumbling, or indeed any other kind of awkwardness." and I'm impressed with how natural that comes out.
Now the Dutch voices on "Ik kan nogal complexe, maar niet gecompliceerde, zinnen zeggen zonder te struikelen, of wat voor foutjes dan ook." (not quite a translation, but the same structure) are clearly robots as soon a you hit the first inter-word pause. Also they make pronunciation errors (wrong vowel in "te"). But, again, the overall inflexion of the sentence sounds right.
These guys are onto something.
I did a load of work on this in 1994 and, clearly, they have some better inflection and frequency models than we did then. Also the pacing for English (and probably Polish) is much better than then. However, it's clear that the phoneme splitting and reconstruction is not always being done correctly. Which probably reflects on the language skills of the people doing this tedious and exacting work. The corpus of sentences being split may also vary quite a bit in size for each language. That will make quite a difference when doing contextual reconstruction.
The real problem is not realistic phono-syntactic synthesis, but the pragmatics of style. It will practically take full-blown AI, or tagged text, to properly render the ever-changing affect implied by characters' content in a novel.
Amazon books are heavily dependent on fiction readers. After several chapters, only the sight-impaired are going to stick with the voices. People are just going to continue sight-reading.
I checked their website, and unless I missed something, they've made a strange oversight: the languages are all European.
Call me crazy, but I suspect this "strange oversight" is what we sometimes call "developing what you have the expertise and resources to develop".
Amazon will miss a huge market in India, Japan, Korea, and of course, China.
Perhaps - one can only hope - Amazon will be able to supply additional resources and expertise?
It was very impressive, but still not a patch on Rutger Hauer...
"I've seen things you people wouldn't believe. Attack ships on fire off the shoulder of Orion. I watched c-beams glitter in the dark near the Tannhäuser Gate. All those moments will be lost in time, like tears in rain. Time to die."
Hello and again welcome to the Aperture Science Computer-Aided Enrichment Centre. We hope your brief detention in the relaxation vault has been a pleasant one. Your specimen has been processed and we are now ready to begin the test proper. Before we start, however, keep in mind that although fun and learning are the primary goals of the enrichment centre activities, serious injuries may occur. For your own safety, and the safety of others, please refrain from GFRXTHGGGHHHNAAAK
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
