Of course, most everything I wrote there is sheer speculation about NVidia's intent for the future . . . but the potential is definitely there to change our concept of what a desktop - hell, even a handheld (eventually) computer can do.
Once this trickles down to the consumer market - it will open up things like true speech recognition (not the crap we call SR now), which could in turn change the way we think about human-machine interfaces.
I'm not so sure about speech recognition. The nifty thing about the GPU is that it is speciality silicon - the procedures that best render 2-d and 3-d graphics have, in the GPU, been rendered in silicon - instead of in sotware. Turns out those procedures are also what many scientists need. GPU's were able to do that - and do it well - but required programs written down to the "bare metal" - programs had to be written without the benefit of a high-level language.
Nvidia said: we're already heading in the direction of an API for our graphics card. let's do it.
One of the big problems in speech recognition is distinguishing speakers from each other and background noise. Whether the speciality processes in silicon on the GPU can be applied to that problem - I think that's up in the air.
What I /know/ it /will/ do: Make speech /synthesis/ much better. Even the best speech synthesis sounds like .... MC Hawking. It's pre-rendered chunks. The speciality processes will allow far better synthesis, and synthesis for difficult spoken languages like Mandarin and Navajo. And Finnish. And maybe with that capability (on-the-fly), someone will figure out how to get it to comprehend speech on the fly as well.
I don't think we will see the Nvidia gpu-as-offload in portable devices soon. AMD bought ATI, and I think they're going to offload their plain CPUs and multi-cpu-core-CPUs to fabless manufacturing model and for the consumer electronics (and handheld computers are sadly considered consumer electronics), I think they're going to put in one package - maybe on one wafer - CPU/GPU/TPM/lotsa RAM - and if they're smart they'll do it on consumer electronics before dropping it into PCs.
I agree on the speech synthesis advancement potential.
The big problem facing speech recognition is complex. There are three primary methods in use today:
One uses a small corpora table and does a real-time analysis for a rough match to sounds in that limited vocabulary list, much like a fuzzy database lookup. It's fairly accurate and can "understand" a wide variety of voices. This is the system used on most modern phone attendant packages. (the ones we all hate so much when we really want a live operator . . .)
Then there is the training system. It's a combined corpora and dictionary lookup. It "learns" a list of phonemes (not just words) from a specific user through "training." On recognition it combines that learned sound list with a dictionary for results. It's the method used in the original ATT/Microsoft Research efforts on their SAPI releases.
Another method still very much in it's infancy has the ability to learn and filter background noise, and can adapt to any users voice very quickly. It's needs in processing power are vast though . . . and that's where I hope Tesla (or something like it) can help.
![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
(no subject)
Date: 2007-06-21 07:06 pm (UTC)(no subject)
Date: 2007-06-21 07:07 pm (UTC)(no subject)
Date: 2007-06-21 07:22 pm (UTC)(no subject)
Date: 2007-06-21 07:25 pm (UTC)(no subject)
Date: 2007-06-21 08:13 pm (UTC)Once this trickles down to the consumer market - it will open up things like true speech recognition (not the crap we call SR now), which could in turn change the way we think about human-machine interfaces.
And more, much more.
(no subject)
Date: 2007-06-21 09:28 pm (UTC)Nvidia said: we're already heading in the direction of an API for our graphics card. let's do it.
One of the big problems in speech recognition is distinguishing speakers from each other and background noise. Whether the speciality processes in silicon on the GPU can be applied to that problem - I think that's up in the air.
What I /know/ it /will/ do: Make speech /synthesis/ much better. Even the best speech synthesis sounds like .... MC Hawking. It's pre-rendered chunks. The speciality processes will allow far better synthesis, and synthesis for difficult spoken languages like Mandarin and Navajo. And Finnish. And maybe with that capability (on-the-fly), someone will figure out how to get it to comprehend speech on the fly as well.
I don't think we will see the Nvidia gpu-as-offload in portable devices soon. AMD bought ATI, and I think they're going to offload their plain CPUs and multi-cpu-core-CPUs to fabless manufacturing model and for the consumer electronics (and handheld computers are sadly considered consumer electronics), I think they're going to put in one package - maybe on one wafer - CPU/GPU/TPM/lotsa RAM - and if they're smart they'll do it on consumer electronics before dropping it into PCs.
(no subject)
Date: 2007-06-21 09:44 pm (UTC)The big problem facing speech recognition is complex. There are three primary methods in use today:
One uses a small corpora table and does a real-time analysis for a rough match to sounds in that limited vocabulary list, much like a fuzzy database lookup. It's fairly accurate and can "understand" a wide variety of voices. This is the system used on most modern phone attendant packages. (the ones we all hate so much when we really want a live operator . . .)
Then there is the training system. It's a combined corpora and dictionary lookup. It "learns" a list of phonemes (not just words) from a specific user through "training." On recognition it combines that learned sound list with a dictionary for results. It's the method used in the original ATT/Microsoft Research efforts on their SAPI releases.
Another method still very much in it's infancy has the ability to learn and filter background noise, and can adapt to any users voice very quickly. It's needs in processing power are vast though . . . and that's where I hope Tesla (or something like it) can help.
(no subject)
Date: 2007-06-23 02:36 pm (UTC)