Slashdot Mirror
Rest In Peas — the Death of Speech Recognition
An anonymous reader writes "Speech recognition accuracy flatlined years ago. It works great for small vocabularies on your cell phone, but basically, computers still can't understand language. Prospects for AI are dimmed, and we seem to need AI for computers to make progress in this area. Time to rewrite the story of the future. From the article: 'The language universe is large, Google's trillion words a mere scrawl on its surface. One estimate puts the number of possible sentences at 10^570. Through constant talking and writing, more of the possibilities of language enter into our possession. But plenty of unanticipated combinations remain, which force speech recognizers into risky guesses. Even where data are lush, picking what's most likely can be a mistake because meaning often pools in a key word or two. Recognition systems, by going with the "best" bet, are prone to interpret the meaning-rich terms as more common but similar-sounding words, draining sense from the sentence.'"
Buffalo buffalo Buffalo buffalo buffalo, buffalo Buffalo buffalo.
> meaning often pools in a key word or two
It's true.
My own hearing is not great. I often miss just a word or two in a sentence. But they are often key words, and missing them leaves the sentence meaningless. If I counted the words I understand correctly I'd probably have a 95% success rate. But if I counted the sentences I understand correctly, I'd be around 80%. So I get by, but I tend to annoy people when I ask for repeats over one missed word.
I hardly type anything in to my HTC Incredible. Google's voice recognition, which is enabled on every textbox works just about perfectly.
Seriously, get an Android phone, try out the speech recognition text entry, and then tell me speech recognition is dead.
That summary was written with speech recognition software?
Natural language processing *is* AI. And high accuracy speech recognition requires natural language processing if we expect to have accuracy rates approaching that of a human. Humans hear words partially or incorrectly all the time. We fill in the gaps from context, and we correct if the course of the conversation reveals that the original interpretation is wrong. Expecting computers to do better, when half the time the problem is the speaker, not the listener, means you need it to be able to make the same corrections from limited information on the fly, and after the fact that a human brain makes.
$_ = "wftedskaebjgdpjgidbsmnjgcdwatb"; tr/a-z/oh, turtleneck Phrase Jar!/; print
It only flatlined because nobody tried to write speech recognition software in perl*.
*Disclaimer: Poster is not responsible for attempts resulting in unintended AI development and/or end of the world scenarios brought on by such an irresponsible endeavor.
Motorcycles, Robots, Space Gossip and More!
Even humans mishear speech.
"'Scuse me while I kiss this guy"
That misheard lyric is so common that there's a book about misheard lyrics with that as the title.
--
BMO
One estimate puts the number of possible sentences at 10^570
What a completely useless metric. It makes sense to examine the context and meaning of speech in order to accurately transcribe words, but the number of possible sentences doesn't seem to accurately describe the problem here...
"Before criticizing someone, first walk a mile in his shoes. Then, you'll be a mile away... and you'll have his shoes."
I've been using VR in Win7 for a few weeks now. I can honestly say that after a few trainings, I'm near 100% accuracy. Which is 15% better than my typing!
I doubt it is completely dead. I have yet to hear it from the researchers working on AI. I work in affective computing, so I am thinking that it is possible that the missing component could be emotion or another way to increase the understanding and ability of computers to learn. In addition, even if it is not possible to increase speech recognition capabilities in this model of computing, in another model of computing this and more would be possible. I am not believing it until I hear it from researchers who have tried most possible options for improvement.
Speech recognition mechanisms/algorithms are not entirely
the problem. What needs to back them up is called a "world
model," and, as the name implies, this can be large and open
ended. Humans being able to correct spoken/heard errors
on the fly is because of having an underlying world model.
Would that I had mod points today.
The above is a valid English sentence and a poignant example of how difficult it is to parse language without knowledge of semantics.
Having said that, Dragon works fairly well, provided you modulate your speech.
If you want a laugh with Dragon, turn away from the screen and talk normally, then look at what it has transcribed..
http://slashdot.org/~GuyFawkes/journal
Futurists should really learn what the word "plateau" means. The death of any given technical progression, particularly one that deals with information procesing, tends to be announced early and often, right up to the point where progress becomes meaningful again and then all of a sudden everyone saw it coming, and oh by the way where's my flying car?
I see a lot of claims, but not much evidence. If we're going to use perceptions and anecdotes as evidence, my impression is that speech recognition has always been considered vaguely stalled. In 2000, people didn't think much progress had been made since 1991 besides some commercialization of stuff academia already knew how to do. In 2010, this guy doesn't think much progress has been made since 2001 besides some commercialization of stuff academia already knew how to do. Yet I think some progress has been made over the past 20 years. There just haven't been any breakthroughs, which is maybe what he's expecting, given his vague suggestion that "AI", a pretty vague concept, is our hope.
I'm also skeptical that accuracy has flatlined, though it's possible that's true in some areas. My impression is that multi-speaker recognition, use of large corpora to improve accuracy, and use of language modeling to improve accuracy, have all improved over the past 10 years. Of course, not all improvements go everywhere: the speech recognition running in real-time on a mobile ARM processor is not using every possible state-of-the-art technique. The advance there is that you can run speech recognition in real-time on a mobile ARM processor at all, and get performance that was once only possible on pretty hefty workstations.
10 PRINT CHR$(205.5+RND(1)); : GOTO 10
It works great for small vocabularies on your cell phone
No. It doesn't.
It works great for small vocabularies on your cell phone if you happen to live in the same neighbourhood as the developer where "everyone talks this way". For the rest of the world, attempting to talk with a nasal American twang in order to get the phone to understand you, is shit.
Deleted
Blame Startrek for making it look flawless. Speech recognition is just like fusion technology, 20 years away from properly working - just like it has been for the last 20 years.
-RANT- I cant stand voice recognition systems that don't at least give you an option to press a number. Especially when they are out of tune and pick up back ground noises as voice. Please, please, please - always give the option to press a number instead of having to voice everything!!
Years ago I used viavoice on Warp4, and it had a pretty decend recognitation rate ..
Looks like whatever you're using now ain't quite as good.
God invented whiskey so the Irish would not rule the world.
Comment removed based on user account deletion
That dosen't always work. When accents and the command of a language are so poor, you only get a few chances to ask, "Sorry, what did you say?" After asking three times, you either look like an asshole and/or give up and spend the next few minutes nodding and smiling before trying to parse what they said, hoping you get it right.
Which is why we need good speech-recognition and translation software. It's easy to infer the meaning of "come to me give the diagram" because there are at least intelligible words to work with. And no, I'm not being racist -- the situation applies to all cultures and languages.
Didn't IBM a few years ago announce a big five-year-program to crack speech recognition? Whatever came of that?
How hard is it for a computer to understand the sentence: "Tea, Earl Grey, Hot"? That takes care of 90% of the use case scenarios right there. Next is "Computer, initiate auto-destruct sequence" is the next 8%.
This blog post is retarded. The author is correlating a drop in internet news articles about Dragon NaturallySpeaking with a flatlining of speech recognition accuracy rate.
The Slashdot editor Soulskill is retarded for both not realizing this and for not reading the anonymously-submitted blog post (hmm no way it could have been the author) before approving it for the Slashdot front page. The guy is just out for more traffic to his rather pointless tech news commentary blog.
Decline of Slashdot, internet signal-to-noise ratio, get off my lawn, etc.
I'd settle for a grammar checker. From the fine summary:
"Even where data are lush"
A good one would have saved this summary from sounding stupid.
snaaaaaaake!
A few years back I worked for an awesome company that did a IVR (interactive voice recording) systems.
We had voice driven interactive systems that would provide the caller with a variety of different mental health tests (we work a lot with identifying depression, early onset dementia, Alzheimer, and other cognitive issues.
The voice recognition wasn't perfect, but we had a review system that dealt with a "gold standard". I wrote a tool that would allow a human being to identify individual words and to label them. Then we would run a number of different voice recognition systems against the same audio chunk and compare their output to the human version. It effectively allowed us to unit test our changes to the voice recognition software.
Dialing in a voice recognition system is an amazing process. The amount of properties, dictionaries, scripting, and sentence forming engines are mind blowing.
Two of the hardest tests for our system were things like: Count from 1 to 20 alternating between numbers and letters as fast as you can, for example 1-A-2-B-3-C. And list every animal you can think of.
The 1-A-2-B was killer because when people speak quickly, their words merge. You literally start creating the sound of the A while the end of the 1 is still coming. It makes it extremely difficult to identify word breaks and actual words. And if you dial in a system specifically to parse that, you'll wind up with issues parsing slower sentences.
The all animals question had a similar issue, people would slur their words together, and the dictionary was huge. It was even more challenging when one of the studies that was nation wide. We had to deal with phonetic spellings from the north east coast and southern states accents. What was even worse was that there was no sentences. We couldn't count on predictive dictionary work to identify the most likely word out of those that would match the phonetics.
That said, getting voice recognition to work on pre-scripted commands and sentences was pretty easy.
And I can only imagine the process has been improving in the years since. Although we were looking into SMS based options, not for a dislike of IVR, but because our usage studies with children were showing most of them were skipping the voice system and using the key pad anyway. So why bother with IVR if the study's target demographic was the youth.
-Rick
"Most people in the U.S. wouldn't know they live in a tyrannical state if it walked up and grabbed their junk." - MyFirs
People want "human quality" speech recognition.
As if we're ever going to get away from training speech recognition programs when we train listeners every day when we speak. It's just that most people don't look at it as being trained, since we're so used to doing it.
I'm sure you have more trouble understanding someone with a thick Cockney or Scottish accent if you're from the Midwest US. You'd ask that person to repeat a few times, wouldn't you?
To expect speech recognition programs to *not* use training is to expect them to exceed human intelligence. Indeed, it's to expect such programs to be psychic.
--
BMO
Long ago - decades, before Bill Gates was invented, a lot of research went into what would be required for actual voice recognition.
A counterexample was given, about an engineering marvel (of the time) that would recognise when someone said the word "watermelon". For a long time, people in the industry assumed that the path to voice recognition consisted of building more and better watermelon boxes.
Several authors, including Alan Turing himself, argued that actual voice recognition could never be accomplished with a large array of watermelon boxes. Current VR software divides input into a series of hyperplanes, and attempts to build a best match from the classification tree.
THis is the 2010 version of the watermelon box.
Real voice recognition won't be practical until the input is parsed, matched against context, and structured much akin to diagramming a sentence in those old English (or other) classes. In short, matching against a vocabulary is trying to solve an exponential problem with a (large) polynomial engine.
It won't be until the computer actually understands what is said that VR is likely to be practical in a global sense.
As a person who has been building computer systems for 35 years, it bothers me to see a huge body of research done into subjects like these ignored, because someone thinks that none of it applies to PC's.
Don't take life too seriously; it isn't permanent.
Obligatory UF
"Nine times out of ten, starting a fire is not the best way to solve the problem." - my wife
I know it's just an imaginary example of how bad text-to-speech is... but it is realistic and disappointing.
Even an idiot like me knows what Markov chain is. Perhaps the standard voice apps are so entrenched they're not recoding their apps to take advantage of huge leaps in memory capacity compared to when they first started selling.
Cwm, fjord-bank glyphs vext quiz
Considering Google is now offering automatic transcription of all YouTube videos, I'd say they certainly haven't given up on speech recognition yet.
English, I would think is a pretty daunting language for speech recognition, what with a substantial array of homophones, but I wonder if other languages fare better. Maybe Spanish or, say, Japanese would be better since (I'm guessing) there is a closer relation to the written script and the actual sound that it makes.
Once I was a four stone apology. Now I am two separate gorillas.
I have been flamed more than a few times around here for suggesting Computer Science has not got a clue what they are doing when it comes to AI. Philosophy has been at this problem and more for the better part of the last 400+ years (more like a 1,000 years) in a serious way. The stock b.s., I get from the science fiction fan boys is that somehow natural language is a problem that can just be brute forced as if you were trying to figure out the password you forgot to your email account. Good luck with that.
By the way, language "recognition" by a computer is likly the easy part of the problem for AI researchers to crack. It is still not going to yield any real AI, just better cars and toasters.
Living in Chile
When I started on my Ph.D., I started out majoring in AI. One of several reasons I changed to computer architecture (CPU design, etc.) is because I just couldn't stand the broken ways that people were doing stuff. Actually computer vision stuff isn't so bad -- at least there's room for advancement. But the speech recognition state of the art is just awful. I couldn't stand the way they did much of anything in pursuit of human language understanding.
With automatic speech recognition (ASR), the first problem is the MFCCs. (Mel-frequency cepstral coefficients.) What they essentially do is take a fourier transform of a fourier transform of the data. This filters out not only amplitude but also frequency, leaving you only with the relative pattern of frequency. Think of this as analogous to taking a second derivative, where all you get is accelerating, leaving out position and velocity. You lose a LOT of information. Then once the MFCC's are computed, they're divided up into the top 13 (or so) dominant MFCCs, plus the first and second step-wise derivatives, giving you a 39D vector. Then the top N most common ones are tallied, and code-booked, mapping the rest to the nearest codes, leaving you with a relatively small number of codes (maybe a few hundred).
So to start with, the signal processing is half deaf, throwing away most of the information. I get why they do it, because it's speaker independent, but you completely lose some VERY valuable information, like prosodic stress, which would be very useful to help with word segmentation. Instead, they try to guess it from statistical models.
Next, they apply a hidden Markov model (HMM). Instead of inferring phones from the signal, the way they model it is as a sequence of hidden states (the phones) that cause the observations (the codes). This statistical model seems kinda backwards, although it works quite well, when trained properly. To train it, you need a lot of labeled data, where people have taken lots of speech recordings and manually labeled the phonetic segments. What is usually learned is mostly a unigram, where what you know are the a priori probabilities of each phone label (the hidden states), and the posterior probability of each phone given each possible prior phone. Given a sequence of codes, you find the most likely sequence of phones by computing the viterbi path through the HMM.
Honestly, I can't complain too much about the HMM. What I do complain about is the fact that the "cutting edge" is to replace the HMM with a markov random field (just remove the arrows from the HMM), and conditional random fields (which are markov random fields with extra inputs).
My response to using MRFs and CRFs is "big whoop", because all you're doing is replacing the statistical model, which doesn't dramatically improve recognition performance, because they haven't fixed the underlying problem with the signal processing.
Then on top of the phone HMM, they layer ANOTHER HMM on top of it to infer words and word boundaries, based on a highly inaccurate phone sequence.
The main problem with all of this is not that the reseachers are idiots. They're not. The problem is that the people with the funding are totally unwilling to fund anything really interesting or revolutionary. The existing methods "work", so the funding sources figure that we can just make incremental changes to existing technologies. Which is wrong. Unfortunately, any radically new technology would be highly experimental, with a high risk of failure, and would take a long time to develop. No one wants to fund anything that iffy. As a result, all the scientists working in this are spend their time on nothing but boring tweaks of a broken but "proven" reasonably effective technology.
So I don't blame people for the conundrum, but I see no opportunity to do anything interesting, so I just couldn't stand studying it.
Only you talk like you. There is no archive of speech large enough to encompass every speaker of a language except one that has a record of each and every speaker. And it still doesn't solve the teaching problem. The shotgun approach is problematic in many ways, most of all the size of the database and you'd still wind up teaching the speech platform to find what accent you're using, because if you ask most people, they don't have any accents at all.
Actually, I think the solution would be to make personal datasets portable, to standards, so when you go from one device to another, all you need to do is plug in your own dataset (or access it from the network) et voila, instant voice recognition wherever you go by systems designed to use that dataset standard. Sort of like an ODF for speech datasets. This way it's distributed, you don't have a humongously unwieldy database to manage, and it's personalized.
But that requires standards which don't yet exist, because every speech recognition platform reinvents the wheel every single time.
--
BMO
Its just a speed bump on the way to thought recognition, which will be far more useful.
I run: Windows, OS X, Linux, FreeBSD. Just because you have a hammer, doesn't mean everything is a nail.
"she helped my uncle jack off a horse"
No, I won't to use a common dataset to train all software automatically, like VoxForge. What I was saying is that people don't need training to talk to each person they meet. A generic background training works fine, and so it should for computers.
Dilbert RSS feed