I'll admit we haven't measured recently, but the speech-to-phone error rate was around 60% not so long ago. OTOH, speech-to-word is closer to 10% by now. Lexical information helps a lot.
Maybe you could do a phone lattice then decode the lattice to words, but there wouldn't really be a point and it would be extremely slower.
The mathematical models for speaker id and speech recognition (plps, gmms, map adaptation, gmm decimation, etc...) are identical, it's only the way they're used that is different, but the code is essentially common.
The "open source community" also needs data. And it isn't available.
Did you notice that you don't have the data to rebuild the models they use in sphinx 4? That means you can't add all the missing parts in it you need if you want to really be state of the art, like VTLN, MMIE/MPE, SAT, or even a lattice rescoring using 4-grams. Kinda kills the whole "other people can make it better" idea...
ASR is just a tool. ASR as itself isn't interesting, and indeed "computer, switch to bold, type xyzzy, back to normal, type..." sucks.
But now, what do you think about "computer, what is the latitude of Tokyo" and have it google it for you. Or "computer, take the file xyzzy.txt on the memory stick I just plugged in and show me a distribution graph of the first column of values". That's just mount, find, awk and gnuplot in practice, we have all the tools already.
You simply don't do the same things with speech than with a keyboard.
DHCP only implies UDP, and that's rather easy. In fact, I participated in a project where it took ~6 months to get IP/UDP/ARP/DHCP running in VHDL on a FPGA, and let me tell you, it's way harder than in pure software.
Don't try to fix it. Crypto algos design should be left to professionals, and I'm very much not one. I suspect, from memory, that what you're looking for, that is turning a hash into a cypher, is called "HMAC". Google for it:-)
And if you can guess the original value of any of the length-sized blocks you can xor the pad out and decrypt the whole document from this point onwards. As it happens, file headers tend to be predictable.
Also two files encrypted with the same key will use the exact same pad. That's not good, and one of the reasons rc4 is frowned upon nowadays.
Whether it is is for the courts to decide. Meanwhile, the libgcc files that come with the compiler say: In addition to the permissions in the GNU General Public License, the Free Software Foundation gives you unlimited permission to link the compiled version of this file into combinations with other programs, and to distribute those combinations without any restriction coming from the use of this file. (The General Public License restrictions do apply in other respects; for example, they cover modification of the file, and distribution when not linked into a combine executable.)
So the compiler itself can't be a problem. And the libc and libc++ tend not to be a problem either (LGPL).
OG.
Re:Blackjack is a game of skill.
on
MAME To Become GPL?
·
· Score: 3, Insightful
A fair video game version of blackjack would be beatable. What makes you think the video gambling games out there are anywhere near fair? Especially when the people programming them tell you to never, ever play them?
OG.
Re:This would be great news if.......
on
MAME To Become GPL?
·
· Score: 3, Interesting
You think wrong. Other people add the recent games whether we want it or not, and we don't care when other versions add the gambling game, we'll just keep not supporting them because they're not considered video games.
BTW, if you're curious, what makes the difference between video games and gambling games for us is whether skill can make a difference. In gambling games, all is decided by the return ratio the operator set in the configuration, and you can't change the results, whatever you do (modulo bugs in the software if there are any exploitable).
OG.
Re:Then why was it removed from the GPL?
on
MAME To Become GPL?
·
· Score: 3, Informative
The initial reason is that one of the cpu cores added in 0.27 had a gpl-incompatible license.
All header files capitalization issues have been fixed though, and any remaining are considered bugs. There is a stray '\' in 0.62, but it's considered a bug too.
The story in Atlantis doesn't have much to do with Nadia. The technology does, but both are blatantly inspired from Jules Verne, hence the common points. Atlantis is more inspired from Jules Verne in practice, including "A journey to the center of the earth" and "Twenty thousand leages under the sea".
No, what they're trying to do is offload the cache management to the virtual memory manager. With mmap() backed by the partition itself, the VM can read and write the pages transparently w.r.t the ext2 server. With read/write/lseek, you have to do actual memory management. Last time I looked, there was no interface for collaboration between the VM and the servers for cache management.
And this kind of cache management is horribly hard in a monolithic kernel for a start. Look how long 2.4 took before the VM behaviour was considered decent (2.4.16 iirc). A decently fast distributed one is even worse to design.
From a design point of view, it's simple. You have a gps, and some compuiter that will give you some data (i.e., a decription key) when the gps detects that you're at a specific position in space. The really, really hard part is making the device tamperproof.
It has not only to resist to direct attacks trying to get to the data, it also has to deal with jamming of the gps signals, or more specifically putting the device in a faraday cage and sending it signals imitating the gps satellites in the appropriate position. Too bad the article has zero information on their methods.
Oh well, let's hope a followup article by Schneier (who also considers the tamperproofing critical) will be more detailed on the technical side.
That's exactly the reason why the FSF is so adamant about getting copyright assignments for the code they accept in their projects, including a release by the employers of the contributors. Otherwise, you have a non-negligible risk of ending up with this kind of problems...
We use Viavoice for linux, and the accuracy is around 85-90% from 3 meters away in a lab with ~50Db of fan noise. The array rules. That's after training Viavoice, of course, and with a speaker that knows how to over-articulate.
If I were you, I'd suspect the microphone and/or the sound card. You should have better results in the first place.
They don't really have a choice, it's the law. It's called maximizing shareholder value. Enjoy.
OG.
Re:Definition of frequency?
on
Share The Pi!
·
· Score: 1
The point is, nobody ever has proven that Pi contains every possible string of numbers. I'm not even sure it has been proven that all the digits are equi-probable.
Just remove the "to criminals" part. It's for selling lock picks, period. What use is done of them is irrelevant. Whether anybody did anything illegal with them is irrelevant. Welcome to the USA.
Slashdot Mirror
I'll admit we haven't measured recently, but the speech-to-phone error rate was around 60% not so long ago. OTOH, speech-to-word is closer to 10% by now. Lexical information helps a lot.
Maybe you could do a phone lattice then decode the lattice to words, but there wouldn't really be a point and it would be extremely slower.
OG.
The mathematical models for speaker id and speech recognition (plps, gmms, map adaptation, gmm decimation, etc...) are identical, it's only the way they're used that is different, but the code is essentially common.
OG.
The "open source community" also needs data. And it isn't available.
Did you notice that you don't have the data to rebuild the models they use in sphinx 4? That means you can't add all the missing parts in it you need if you want to really be state of the art, like VTLN, MMIE/MPE, SAT, or even a lattice rescoring using 4-grams. Kinda kills the whole "other people can make it better" idea...
OG.
Also know as "broadcast news" is 8.6% WER at 10xRT and 11.8% WER at 1xRT. 18.7% at 4xRT is so last millenium...
OG.
ASR is just a tool. ASR as itself isn't interesting, and indeed "computer, switch to bold, type xyzzy, back to normal, type ..." sucks.
But now, what do you think about "computer, what is the latitude of Tokyo" and have it google it for you. Or "computer, take the file xyzzy.txt on the memory stick I just plugged in and show me a distribution graph of the first column of values". That's just mount, find, awk and gnuplot in practice, we have all the tools already.
You simply don't do the same things with speech than with a keyboard.
OG.
Nobody knows how to actually do it at that point, open source or not.
OG.
DHCP only implies UDP, and that's rather easy. In fact, I participated in a project where it took ~6 months to get IP/UDP/ARP/DHCP running in VHDL on a FPGA, and let me tell you, it's way harder than in pure software.
TCP on the other hand is really complex.
OG.
It indeed won't be there in 10 years. It will rather be 5, at most, maybe even less.
OG.
Don't try to fix it. Crypto algos design should be left to professionals, and I'm very much not one. I suspect, from memory, that what you're looking for, that is turning a hash into a cypher, is called "HMAC". Google for it :-)
OG.
And if you can guess the original value of any of the length-sized blocks you can xor the pad out and decrypt the whole document from this point onwards. As it happens, file headers tend to be predictable.
Also two files encrypted with the same key will use the exact same pad. That's not good, and one of the reasons rc4 is frowned upon nowadays.
XOR is very, very hard to use correctly.
OG.
Whether it is is for the courts to decide. Meanwhile, the libgcc files that come with the compiler say:
In addition to the permissions in the GNU General Public License, the
Free Software Foundation gives you unlimited permission to link the
compiled version of this file into combinations with other programs,
and to distribute those combinations without any restriction coming
from the use of this file. (The General Public License restrictions
do apply in other respects; for example, they cover modification of
the file, and distribution when not linked into a combine
executable.)
So the compiler itself can't be a problem. And the libc and libc++ tend not to be a problem either (LGPL).
OG.
A fair video game version of blackjack would be beatable. What makes you think the video gambling games out there are anywhere near fair? Especially when the people programming them tell you to never, ever play them?
OG.
You think wrong. Other people add the recent games whether we want it or not, and we don't care when other versions add the gambling game, we'll just keep not supporting them because they're not considered video games.
BTW, if you're curious, what makes the difference between video games and gambling games for us is whether skill can make a difference. In gambling games, all is decided by the return ratio the operator set in the configuration, and you can't change the results, whatever you do (modulo bugs in the software if there are any exploitable).
OG.
The initial reason is that one of the cpu cores added in 0.27 had a gpl-incompatible license.
OG.
All header files capitalization issues have been fixed though, and any remaining are considered bugs. There is a stray '\' in 0.62, but it's considered a bug too.
OG.
The story in Atlantis doesn't have much to do with Nadia. The technology does, but both are blatantly inspired from Jules Verne, hence the common points. Atlantis is more inspired from Jules Verne in practice, including "A journey to the center of the earth" and "Twenty thousand leages under the sea".
OG.
No, what they're trying to do is offload the cache management to the virtual memory manager. With mmap() backed by the partition itself, the VM can read and write the pages transparently w.r.t the ext2 server. With read/write/lseek, you have to do actual memory management. Last time I looked, there was no interface for collaboration between the VM and the servers for cache management.
And this kind of cache management is horribly hard in a monolithic kernel for a start. Look how long 2.4 took before the VM behaviour was considered decent (2.4.16 iirc). A decently fast distributed one is even worse to design.
OG.
Yes, no decimal. That's how the meter is defined, using the second as a reference.
OG.
"I'm at the edge of the mousepad, I can't go to the right anymore"
"Well, pick up the mouse and move it to the left of the pad"
"But even picked up the pointer still follow when I move to the left"
"Ah, you need our $2000 desk upgrade then"
OG.
From a design point of view, it's simple. You have a gps, and some compuiter that will give you some data (i.e., a decription key) when the gps detects that you're at a specific position in space. The really, really hard part is making the device tamperproof.
It has not only to resist to direct attacks trying to get to the data, it also has to deal with jamming of the gps signals, or more specifically putting the device in a faraday cage and sending it signals imitating the gps satellites in the appropriate position. Too bad the article has zero information on their methods.
Oh well, let's hope a followup article by Schneier (who also considers the tamperproofing critical) will be more detailed on the technical side.
OG.
That's exactly the reason why the FSF is so adamant about getting copyright assignments for the code they accept in their projects, including a release by the employers of the contributors. Otherwise, you have a non-negligible risk of ending up with this kind of problems...
OG.
We use Viavoice for linux, and the accuracy is around 85-90% from 3 meters away in a lab with ~50Db of fan noise. The array rules. That's after training Viavoice, of course, and with a speaker that knows how to over-articulate.
If I were you, I'd suspect the microphone and/or the sound card. You should have better results in the first place.
OG.
They don't really have a choice, it's the law. It's called maximizing shareholder value. Enjoy.
OG.
The point is, nobody ever has proven that Pi contains every possible string of numbers. I'm not even sure it has been proven that all the digits are equi-probable.
OG.
Just remove the "to criminals" part. It's for selling lock picks, period. What use is done of them is irrelevant. Whether anybody did anything illegal with them is irrelevant. Welcome to the USA.
OG.