12 ms·

Google has developed speech-recognition technology that actually works.

101 points by pelf 16y ago

MatthewPhillips 16y ago
I wonder at what point Apple is going to have to start building these kinds of technologies. Purchasing Siri was supposed to be a step in that direction but nothing has come of it yet. I worry that it's not in Apple's DNA to build products they can't directly charge for. Apple doesn't do freemium. The hope would be that third parties would pick up the slack. However I'm not sure that startups can match Google in bigdata. So how will Apple catch up in voice recognition? In mapping (which also uses Android data to improve things like traffic and rerouting)?
- jokermatt999 16y ago
  I honestly can't see Apple doing speech recognition. It's good, but it's only 99% good, and Apple are known for being perfectionists. I'm not sure about mapping, it also doesn't seem like an Apple product to me, but I can't really articulate why.
  - mdemare 16y ago
    That is a cliche, and an incorrect one. Take autocorrect for instance - hardly a perfect product. (http://damnyouautocorrect.com/ http://damnyouautocorrect.com/)
  - cdavid 16y ago
    You sometimes see one or two Apple people at speech conferences, at least it was true 2-3 years ago. I cannot imagine that not to be the case anymore.
- pkulak 16y ago
  Apple will make great speech recognition, maps and email when Google makes a great UI.
kajecounterhack 16y ago
Does anyone happen to know if there are significant companies in the speech-recognition space besides Dragon and Google?
- CWuestefeld 16y ago
  One assumes that the NSA is doing a lot.
  - hyperbovine 16y ago
    Yes: http://en.wikipedia.org/wiki/Room_641A http://en.wikipedia.org/wiki/Room_641A I doubt they're inspecting all that traffic just to watch the pretty patterns.
    - JonnieCache 16y ago
      Surely you mean https://secure.wikimedia.org/wikipedia/en/wiki/Room_641A https://secure.wikimedia.org/wikipedia/en/wiki/Room_641A ?
  - asuth 16y ago
    When I took a speech rec class last semester, we had a guy from BBN (subsidiary of Raytheon) give a talk about large-scale, extremely fast audio transcription. As in, systems that could process audio 30-40,000 times faster than real-time. They traded off recognition accuracy to get this, so their accuracy was around 50-60%, as I recall. I asked why something like that would be useful, and he said if you're looking at a lot of data (which I heard: eavesdropping on an entire telephone network) then all you need is a general idea of what people are saying and a few keywords before you can zoom in on specific clips for more thorough analysis. So yes, I'm sure NSA is interested.
    - Pooter 16y ago
      Yeah. People from Fort Meade are involved in a lot of NLP research, both in terms of primary research (writing their own systems) and funding research and open technology evaluations. Video surveillance systems are getting a lot of attention now too.
orijing 16y ago
Sorry to be a killjoy, but the premise of this article--that Google developed the speech-recognition technology--hurts my feelings (to say the least) and underestimates the contributions of the NLP community. Speech recognition, like machine translation, is academic in origin, and much of the work is still carried out in academia. For example, Google did not "invent" machine translation. No, Google Translate is an adapted version of academic systems. Perhaps the phrase tables are sharded and so is the language model, but the general algorithms are the same. Sure, one of Berkeley's NLP grads is working there, but it's basically an adapted version of what's available. They publish papers like "Stupid Backoff" [1], but that makes them as much a contributor as any other member of the NLP community. Speech recognition is the same thing. Google is the company that takes existing research and adapts it. To claim that Google developed the speech recognition technology is to discredit the contributions of everyone else in the NLP community. Google has been generous at funding NLP research at the university level. Do you consider the results of those research "Google"'s? Ultimately, the main difference is that Google has magnitudes more data and the physical capacity to handle that, not that it solved some systems or architectural bottleneck that has been limiting us. Someone once said that all you need is a crappy model and great data to build a good ML-based algorithm... [1] http://acl.ldc.upenn.edu/D/D07/D07-1090.pdf http://acl.ldc.upenn.edu/D/D07/D07-1090.pdf
- thezilch 16y ago
  Personally, I read it more as Google _applied_ -- developed useful software -- speech-recognition technology that works. I can recall testing countless softwares like Dragon Talk (mentioned in the article) for assisting my writing of school papers, to no avail. These softwares were sluggish and woefully inaccurate -- this may no longer be the case, as I haven't given them a second chance since. I don't think the article disparages past and present work from the NLP community, but you have to expect Google to catch the limelight for the application of such technology. I don't really see anyone else attempting what Google is attempting at scale. Why wouldn't Google be praised (or an "owner") for being a generous benefactor to a community? Would you not consider research at Googleplex (or any of the 100s of satellite campuses) to be Google's?
  - 16y ago
IDisposableHero 16y ago
It got "how much wood would a woodchuck chuck if a wouldchuck would chuck wood" right for me. I am impressed now. But it couldn't handle "Ph'nglui mglw'nafh Cthulhu R'lyeh wgah'nagl fhtagn" Maybe I'm not pronouncing it quite right.
- StavrosK 16y ago
  Are you sure you pronounced all the schwas?
- ppod 16y ago
  "how much wood would a woodchuck chuck if a wouldchuck would chuck wood" That isn't really the best example, considering how these systems tend to work. If their system has a giant bank of text it's using to predict the statistical likelihood of your next word, once you've said "how much wood ..", 'would' is a very high probability candidate for the next word, and once you get to 'woodchuck' the rest is statistically almost inevitable. A better 'difficult' test would be something along the lines of 'colorless green ideas sleep furiously', although we can't actually use that one since that example is so famous it would likely turn up in a web-derived corpus many times.
  - Cyndre 16y ago
    How is that even a test for it? If you input garbage into a system like that your going to get garbage out. And for fun, walk up to a stranger and say something like that and see if the stranger can provide useful information.
  - IDisposableHero 15y ago
    Yes, it kind of failed on "movie times for 'source code'". That's a sentence that's hard to parse without context ;)
martythemaniak 16y ago
"It even works if you've got an accent." I have an indeterminate accent and my voice is on the low side of Bass so I trip up google voice pretty badly - it's mostly useless for me. FWIW, Rockband also can't make sense of what I say. I do wonder when it'll be good enough to understand me.
- sagarm 16y ago
  I hate to break it to you, but Rockband doesn't score you based on what you say. It just scores you based whether you're hitting the right notes. :P
antiterra 16y ago
It seems even strange to me that the article makes no attempt to survey other current examples of speech-recognition technology in order to support the unsaid implication of their lede. That is: "speech-recognition technology developed by those other than Google does not work." I just downloaded the Bing app for iPad last night, and noticed it has a pretty decent speech-recognition engine from a company they acquired a few years back: TellMe. I tried all the examples given in the Slate article, and they were recognized just fine. This makes me curious, are there a number of current-generation speech-recognition technologies that work at the level of Google's? I should note that I didn't receive the desired behavior once my speech was recognized. When I asked the math question, a link to Wolfram Alpha was given, but I would have to click that link to get the answer. I had to go to maps to get any kind of relevant answer for "Directions to McDonalds," and I had to just say McDonalds and then click Directions to get the actual information. This failing appears to be a trait of the iOS app itself. Hand typing the math query into Bing on a proper browser did give me Google Calculator style results.
- ezy 16y ago
  That's because it's Google PR (http://www.paulgraham.com/submarine.html http://www.paulgraham.com/submarine.html). Nuance Communications[1] (who makes Dragon and N other speech rec systems), is really the leader in this space. They've done server-based recognition solutions ~forever, and is the solution Siri uses for their app, for example. [1] Note: this is not the same as the Nuance Mike Cohen started, but a conglomeration of speech companies purchased over the the years which inherited the name. EDIT: Added some more detail
  - Tycho 16y ago
    Yeah, but... i've tried Google Voice Search and it works incredibly well; whereas Dragon on the iPad did not work well enough to actually use for its intended purpose (although it was fairly impressive - i had low expectations).
JesseAldridge 16y ago
I tried it up. It doesn't work went well services dept. Maybe this is my voice doesn't give me a couple of words wrong in all 50 states. Did welcome center for dental bar kansas city missouri. ---- I tried it out. It doesn't work quite as well as the article suggests. Maybe it's just my voice, but it gets at least a couple of words wrong in almost every statement. "It", "welcome", and "thoughts", for example, are consistently misheard.
- jerf 16y ago
  You can really see the "search term bias" in that snippet.
  - nostrademons 16y ago
    I've found the Android speech recognition works great for voice search and sucks for everything else. I've written a HN post or two with it; it either needed much hand editing or an outright translation to be intelligible.
  - orijing 16y ago
    Very interesting observation! It didn't immediately seem obvious to me why there were references to "welcome center", "dental", etc.
- georgemcbay 16y ago
  I've got one years worth of transcribed Google Voice VMs that back you up on this. The way Google does "best fit" matching of words sometimes results in messages that are just hilariously mistranscribed. You can usually get the gist of what the person actually said, which makes the transcription useful, but once in a while the message just diverts off into total nonsense. Here's a recent one: Hey George, it's me does not seem to be working and it's going, but it doesn't seem to be having issue tonight and I was just wondering if that was. If there's anything that we did like If we're not getting any call. Walker, I don't if it's just been forever and it's not Brian anything okay if you get this call. Bye.
- abraham 16y ago
PonyGumbo 16y ago
Hopefully they'll be able to apply this to Google Voice. The voicemail transcriptions are almost always hilariously wrong.
- georgemcbay 16y ago
  Google Voice actually uses the same technology and same dataset, AFAIK, which is why I was so confused by reading this article. Google's stuff is pretty good some of the time, but they've hardly solved this problem to the degree the article suggests (as anyone who has actually used this for more than 5 minutes could tell you).
  - asuth 16y ago
    long-form transcription is a pretty different problem for language models than parsing search queries. There's lots of audio-processing overlap sure, but parsing a voicemail definitely has different, harder challenges.
jodrellblank 16y ago
What are the chances that Google can pick up enough from our voices to biometrically identify us from a crowd in future? E.g. On a Google phone conversation or when licensed to a surveillance company.
micah63 16y ago
Whatever anyone wants to say negatively about this article, it's bang on. I just got my first android phone, it has 2.2 and the little microphone is my new best friend. I talk out todos, write emails, texts, search youtube, search everything, it blows my mind. The only thing it's not good at is uncommon names and places (like my name "Micah" and "Quebec"). Rock on Google.

101 points by pelf 16y ago

I wonder at what point Apple is going to have to start building these kinds of technologies. Purchasing Siri was supposed to be a step in that direction but nothing has come of it yet. I worry that it's not in Apple's DNA to build products they can't directly charge for. Apple doesn't do freemium. The hope would be that third parties would pick up the slack. However I'm not sure that startups can match Google in bigdata. So how will Apple catch up in voice recognition? In mapping (which also uses Android data to improve things like traffic and rerouting)?

jokermatt999 16y ago
I honestly can't see Apple doing speech recognition. It's good, but it's only 99% good, and Apple are known for being perfectionists. I'm not sure about mapping, it also doesn't seem like an Apple product to me, but I can't really articulate why.
- mdemare 16y ago
  That is a cliche, and an incorrect one. Take autocorrect for instance - hardly a perfect product. (http://damnyouautocorrect.com/ http://damnyouautocorrect.com/)
- cdavid 16y ago
  You sometimes see one or two Apple people at speech conferences, at least it was true 2-3 years ago. I cannot imagine that not to be the case anymore.
pkulak 16y ago
Apple will make great speech recognition, maps and email when Google makes a great UI.

kajecounterhack 16y ago

Does anyone happen to know if there are significant companies in the speech-recognition space besides Dragon and Google?

CWuestefeld 16y ago
One assumes that the NSA is doing a lot.
- hyperbovine 16y ago
  Yes: http://en.wikipedia.org/wiki/Room_641A http://en.wikipedia.org/wiki/Room_641A I doubt they're inspecting all that traffic just to watch the pretty patterns.
  - JonnieCache 16y ago
    Surely you mean https://secure.wikimedia.org/wikipedia/en/wiki/Room_641A https://secure.wikimedia.org/wikipedia/en/wiki/Room_641A ?
- asuth 16y ago
  When I took a speech rec class last semester, we had a guy from BBN (subsidiary of Raytheon) give a talk about large-scale, extremely fast audio transcription. As in, systems that could process audio 30-40,000 times faster than real-time. They traded off recognition accuracy to get this, so their accuracy was around 50-60%, as I recall. I asked why something like that would be useful, and he said if you're looking at a lot of data (which I heard: eavesdropping on an entire telephone network) then all you need is a general idea of what people are saying and a few keywords before you can zoom in on specific clips for more thorough analysis. So yes, I'm sure NSA is interested.
  - Pooter 16y ago
    Yeah. People from Fort Meade are involved in a lot of NLP research, both in terms of primary research (writing their own systems) and funding research and open technology evaluations. Video surveillance systems are getting a lot of attention now too.

orijing 16y ago

Sorry to be a killjoy, but the premise of this article--that Google developed the speech-recognition technology--hurts my feelings (to say the least) and underestimates the contributions of the NLP community. Speech recognition, like machine translation, is academic in origin, and much of the work is still carried out in academia. For example, Google did not "invent" machine translation. No, Google Translate is an adapted version of academic systems. Perhaps the phrase tables are sharded and so is the language model, but the general algorithms are the same. Sure, one of Berkeley's NLP grads is working there, but it's basically an adapted version of what's available. They publish papers like "Stupid Backoff" [1], but that makes them as much a contributor as any other member of the NLP community. Speech recognition is the same thing. Google is the company that takes existing research and adapts it. To claim that Google developed the speech recognition technology is to discredit the contributions of everyone else in the NLP community. Google has been generous at funding NLP research at the university level. Do you consider the results of those research "Google"'s? Ultimately, the main difference is that Google has magnitudes more data and the physical capacity to handle that, not that it solved some systems or architectural bottleneck that has been limiting us. Someone once said that all you need is a crappy model and great data to build a good ML-based algorithm... [1] http://acl.ldc.upenn.edu/D/D07/D07-1090.pdf http://acl.ldc.upenn.edu/D/D07/D07-1090.pdf

thezilch 16y ago
Personally, I read it more as Google _applied_ -- developed useful software -- speech-recognition technology that works. I can recall testing countless softwares like Dragon Talk (mentioned in the article) for assisting my writing of school papers, to no avail. These softwares were sluggish and woefully inaccurate -- this may no longer be the case, as I haven't given them a second chance since. I don't think the article disparages past and present work from the NLP community, but you have to expect Google to catch the limelight for the application of such technology. I don't really see anyone else attempting what Google is attempting at scale. Why wouldn't Google be praised (or an "owner") for being a generous benefactor to a community? Would you not consider research at Googleplex (or any of the 100s of satellite campuses) to be Google's?
- 16y ago

IDisposableHero 16y ago

It got "how much wood would a woodchuck chuck if a wouldchuck would chuck wood" right for me. I am impressed now. But it couldn't handle "Ph'nglui mglw'nafh Cthulhu R'lyeh wgah'nagl fhtagn" Maybe I'm not pronouncing it quite right.

StavrosK 16y ago
Are you sure you pronounced all the schwas?
ppod 16y ago
"how much wood would a woodchuck chuck if a wouldchuck would chuck wood" That isn't really the best example, considering how these systems tend to work. If their system has a giant bank of text it's using to predict the statistical likelihood of your next word, once you've said "how much wood ..", 'would' is a very high probability candidate for the next word, and once you get to 'woodchuck' the rest is statistically almost inevitable. A better 'difficult' test would be something along the lines of 'colorless green ideas sleep furiously', although we can't actually use that one since that example is so famous it would likely turn up in a web-derived corpus many times.
- Cyndre 16y ago
  How is that even a test for it? If you input garbage into a system like that your going to get garbage out. And for fun, walk up to a stranger and say something like that and see if the stranger can provide useful information.
- IDisposableHero 15y ago
  Yes, it kind of failed on "movie times for 'source code'". That's a sentence that's hard to parse without context ;)

martythemaniak 16y ago

"It even works if you've got an accent." I have an indeterminate accent and my voice is on the low side of Bass so I trip up google voice pretty badly - it's mostly useless for me. FWIW, Rockband also can't make sense of what I say. I do wonder when it'll be good enough to understand me.

sagarm 16y ago
I hate to break it to you, but Rockband doesn't score you based on what you say. It just scores you based whether you're hitting the right notes. :P

antiterra 16y ago

It seems even strange to me that the article makes no attempt to survey other current examples of speech-recognition technology in order to support the unsaid implication of their lede. That is: "speech-recognition technology developed by those other than Google does not work." I just downloaded the Bing app for iPad last night, and noticed it has a pretty decent speech-recognition engine from a company they acquired a few years back: TellMe. I tried all the examples given in the Slate article, and they were recognized just fine. This makes me curious, are there a number of current-generation speech-recognition technologies that work at the level of Google's? I should note that I didn't receive the desired behavior once my speech was recognized. When I asked the math question, a link to Wolfram Alpha was given, but I would have to click that link to get the answer. I had to go to maps to get any kind of relevant answer for "Directions to McDonalds," and I had to just say McDonalds and then click Directions to get the actual information. This failing appears to be a trait of the iOS app itself. Hand typing the math query into Bing on a proper browser did give me Google Calculator style results.

ezy 16y ago
That's because it's Google PR (http://www.paulgraham.com/submarine.html http://www.paulgraham.com/submarine.html). Nuance Communications[1] (who makes Dragon and N other speech rec systems), is really the leader in this space. They've done server-based recognition solutions ~forever, and is the solution Siri uses for their app, for example. [1] Note: this is not the same as the Nuance Mike Cohen started, but a conglomeration of speech companies purchased over the the years which inherited the name. EDIT: Added some more detail
- Tycho 16y ago
  Yeah, but... i've tried Google Voice Search and it works incredibly well; whereas Dragon on the iPad did not work well enough to actually use for its intended purpose (although it was fairly impressive - i had low expectations).

JesseAldridge 16y ago

I tried it up. It doesn't work went well services dept. Maybe this is my voice doesn't give me a couple of words wrong in all 50 states. Did welcome center for dental bar kansas city missouri. ---- I tried it out. It doesn't work quite as well as the article suggests. Maybe it's just my voice, but it gets at least a couple of words wrong in almost every statement. "It", "welcome", and "thoughts", for example, are consistently misheard.

jerf 16y ago
You can really see the "search term bias" in that snippet.
- nostrademons 16y ago
  I've found the Android speech recognition works great for voice search and sucks for everything else. I've written a HN post or two with it; it either needed much hand editing or an outright translation to be intelligible.
- orijing 16y ago
  Very interesting observation! It didn't immediately seem obvious to me why there were references to "welcome center", "dental", etc.
georgemcbay 16y ago
I've got one years worth of transcribed Google Voice VMs that back you up on this. The way Google does "best fit" matching of words sometimes results in messages that are just hilariously mistranscribed. You can usually get the gist of what the person actually said, which makes the transcription useful, but once in a while the message just diverts off into total nonsense. Here's a recent one: Hey George, it's me does not seem to be working and it's going, but it doesn't seem to be having issue tonight and I was just wondering if that was. If there's anything that we did like If we're not getting any call. Walker, I don't if it's just been forever and it's not Brian anything okay if you get this call. Bye.
abraham 16y ago

PonyGumbo 16y ago

Hopefully they'll be able to apply this to Google Voice. The voicemail transcriptions are almost always hilariously wrong.

georgemcbay 16y ago
Google Voice actually uses the same technology and same dataset, AFAIK, which is why I was so confused by reading this article. Google's stuff is pretty good some of the time, but they've hardly solved this problem to the degree the article suggests (as anyone who has actually used this for more than 5 minutes could tell you).
- asuth 16y ago
  long-form transcription is a pretty different problem for language models than parsing search queries. There's lots of audio-processing overlap sure, but parsing a voicemail definitely has different, harder challenges.

jodrellblank 16y ago

What are the chances that Google can pick up enough from our voices to biometrically identify us from a crowd in future? E.g. On a Google phone conversation or when licensed to a surveillance company.

micah63 16y ago

Whatever anyone wants to say negatively about this article, it's bang on. I just got my first android phone, it has 2.2 and the little microphone is my new best friend. I talk out todos, write emails, texts, search youtube, search everything, it blows my mind. The only thing it's not good at is uncommon names and places (like my name "Micah" and "Quebec"). Rock on Google.