Want to read Slashdot from your mobile device? Point it at m.slashdot.org and keep reading!


Forgot your password?
AI Input Devices Technology

Siri, Cortana and Google Have Nothing On SoundHound's Speech Recognition 235

MojoKid writes: Your digital voice assistant app is incompetent. Yes, Siri can give you a list of Italian restaurants in the area, Cortana will happily look up the weather, and Google Now will send a text message, if you ask it to. But compared to Hound, the newest voice search app on the block, all three of the aforementioned assistants might as well be bumbling idiots trying to outwit a fast talking rocket scientist. At its core, Hound is the same type of app — you bark commands or ask questions about any number of topics and it responds intelligently. And quickly. What's different about Hound compared to Siri, Cortana, and Google Now is that it's freakishly fast and understands complex queries that would have the others hunched in the fetal position, thumb in mouth. Check out the demo. It's pretty impressive.
This discussion has been archived. No new comments can be posted.

Siri, Cortana and Google Have Nothing On SoundHound's Speech Recognition

Comments Filter:
  • by Overzeetop ( 214511 ) on Thursday June 04, 2015 @12:39PM (#49839925) Journal

    Or does it just stare at you stupidly because using ways to give you directions means nothing if it doesn't recognize the homophone.

    • Re: (Score:2, Funny)

      by sycodon ( 149926 )

      "What is the population of capital of the country in which Space Needle is located?"

      Hound correctly surmises that he's asking for the population of Washington, DC...

      The Space Needle is in Seattle.

    • by goombah99 ( 560566 ) on Thursday June 04, 2015 @01:34PM (#49840397)

      Holy crap the video is impressive. It clearly parses phrased and dependent logical statements like " what is the population of the capitol of the country in which the space needle is located. " It alos parsed paragraph long multi-part questions. I was floored.

      As for homophones, how do you (human) recognize them. Well you parse the logical context. If you are doing single word dictation homophones will always be a problem but for queries there's context. And the demo shows this thing can handle some staggering conditional contexts and long phrases. So I would guess that if your query is not ambiguous in the use of the word Waze, then this thing is approachi8ng a level where it will indeed get the right homophone.

      • The challenge is the non-standard homophones. As the smart-ass AC showed, Waze is not a typically recognized homophone (of anything) because it's not a word. Recognition works great with core speech, but anything specialized usually gets mangled. Try "what is the size of a double you twelve by fifty-three." The answer is, of course, is 12 inches by 10 inches. It may be one of the most common sections used in building construction. I'm not sure it would help even if you prefaced it with an "ay eye ess sea"

        • by ceoyoyo ( 59147 )

          To be fair, your query doesn't make any sense to anyone who isn't familiar with an industry where they might run into a w12 (which appears to be a steel I-beam, so that would be fairly heavy construction). If you trained a speech recognition program with construction terms it wouldn't have a problem with that example.

    • It seems like that problem in solved these days. I have no problem saying "Ok google, open waze" and it does the right thing.

    • You do realize that Google Now will happily open Waze if you say "open Waze app"? Give it some context and it knows exactly what to do.

      That said, I agree that s statement beginning " open ..." could automatically be interpreted as meaning an app, but there may be reluctance I do that in case it interferes with future expansion into the internet of things, e.g. "open the curtains", or " open the garage ".

    • Siri launches Waze. Is this an old problem, or a problem on other platforms?

    • by AmiMoJo ( 196126 )

      I just tried saying "launch Waze" to my OnePlus One running Lollipop (5.0.2) and the stock Google Now launcher. The text appeared as "launch ways" for a fraction of a second and then corrected to "launch Waze", and sure enough the Waze app opened.

      It works fine if it has context. Saying "launch" implies I want an app to open. How did you phrase your request?

    • by afidel ( 530433 )

      I used Tasker + Tasker Now to fix that and so many other problems with Google Now. Basically if you tell it to Open Waze Google will correctly use Waze instead of ways, but it won't do anyting, so I added a Tasker Now rule for "Open Waze" that launches the app.

  • Sure this isn't some Baidu thing?

    • Re: (Score:3, Funny)

      by Anonymous Coward

      Nah, people actually use Baidu services. Nobody uses this crap, hence the Slashvertisement. That's what Siri told me anyway. Cortana just giggled.

  • Siri, Cortana and Google are pretty bad compared to the mobile app of "Dragon NaturallySpeaking". Nuance has been the king of voice recognition for both consumer and military use. I doubt soundhound can beat them. If they do, they are in line for some hefty contracts.

    • I know a dictation service that may be out of business if Soundhound's engine beats Dragon for accuracy.
    • by Krojack ( 575051 )

      I try Dragon on and off all the time.. It's a major battery drain compared to Google Now.

      I still keep going back and installing Dragon every few months in hopes this improves though.

    • by lsllll ( 830002 )
      You're missing the point of the app. It's not merely a speech recognition piece of software as much as it is a search engine that understands your plain sentences and is able to find exactly the answer you're after. Hell, show me a web site that I can type any of the sentences in the demo in a search box and have it come up with the answers as quickly as in the demo.
  • I tried sending a text with Google's voice engine last week just to try it out. It did a very good job of taking my dictation to text, then it asked if I wanted to send. I said yes. It spelled out yes in it's little window, then asked again, I said yes again, I tried other words, it also recognized those words, and every time asked me if I wanted to send, while recognizing the words. I finally reached over and hit the send button.

    • by Overzeetop ( 214511 ) on Thursday June 04, 2015 @01:52PM (#49840515) Journal

      It also only works with casual conversation.

      I tried replying to a work text with something like "It's okay to use a W12x14 in place of the C section. Just make sure that it's AISC A992 grade 50" What came out was unusable, while "yo, bitch, put the dinner on the table I'll be home in 5" was transcribed verbatim. Thank goodness I had the same problem with voice send or I would have been picking up McDonalds on my way to sleep with the dog.

      Actually, it really needs to automatically read it back to you, otherwise you have to read what it typed - and that defeats the purpose of being voice activated if you're driving.

  • Holy shit (Score:5, Funny)

    by rebelwarlock ( 1319465 ) on Thursday June 04, 2015 @12:50PM (#49840029)
    Could you suck SoundHound's cock a little harder? This is the most shameless bullshit I've seen all day, and I just watched Kayne West talk for 30 seconds.
    • Re: (Score:2, Insightful)

      by Anonymous Coward

      This is a dice holdings property you are referring to... Slashdot is here for pushing certain political buttons (to keep the readership "engaged") and for advertising to this "engaged" readership (to make money). Slashvertising will only get more aggressive as the readership declines in an attempt to make up falling revenues.

    • by flopsquad ( 3518045 ) on Thursday June 04, 2015 @01:08PM (#49840175)
      It's a slashvertisement, written by a brogrammer, wrapped in a handjob.
    • Re:Holy shit (Score:5, Interesting)

      by Bosconian ( 158140 ) on Thursday June 04, 2015 @01:25PM (#49840273) Journal

      The only thing MojoKid (1002251) wrote for this submission was "Check out the demo. It's pretty impressive," while the rest was plagiarized from the "Hothardware" article written by Paul Lilly, who does seems to be breathlessly impressed by an internal demo of an unreviewed application.

      I'm going to call this a formatting error and a sad omission of credit, because I refuse to believe that someone would shamelessly lift words that they hadn't written and posit them as their own. Maybe it's the editors' fault. In either case, it's sloppy posting and comes off as skeezy no matter what the excuse might be.

      Hell, just submit the rest of the article next time - why bother linking to a source or crediting an original author?

  • by sideslash ( 1865434 ) on Thursday June 04, 2015 @12:51PM (#49840047)
    1. This demo was likely created by an engineer or sales person with SoundHound. More impressive would be a demo by a third party journalist or reviewer without a vested interest.
    2. The impressive speed probably won't scale to the millions of simultaneous users Siri, Google Now, and Cortana support (assuming audio is processed in the cloud, which I admittedly don't know for sure).
    3. Obviously the demo uses phrases that work. I guarantee you an ordinary person will often get "Sorry, I didn't understand the question" or whatever SoundHound's equivalent is.
    4. While it sounds impressive at first blush, nobody really cares how many days it is between next Tuesday and Christmas of 2025. And that happens to be not only useless, but also pretty easy to special-case in your expert system / AI logic. So how about a demo that answers the question: "How can you make a mushroom omelette without soggy mushrooms?"
    • by TWX ( 665546 ) on Thursday June 04, 2015 @12:57PM (#49840095)
      It's also good to be skeptical if this thing doesn't do all of the work on the handheld device and simply send the parsed text to the search engine or other central server to retrieve only the relevant information.

      I mean, c'mon already! I had Dragon running on a friggin' Macintosh LCII in elementary school! That thing was running System 7.1 on a Motorola 68030 with 4MB RAM. Why cant my multi-Gigahertz smartphone with 64GB storage and 4GB RAM do the basic speech-to-text locally that a 25 year old Macintosh can?
    • by phantomfive ( 622387 ) on Thursday June 04, 2015 @01:11PM (#49840181) Journal
      BSD FORTUNE #42!
      "Any sufficiently advanced technology is indistinguishable from a rigged demo."
    • 1) Not likely, the video is on SoundHound's YouTube channel - they're not hiding anything
      2) I honestly don't know how they're processing it that fast... it likely isn't cloud based otherwise there would be a delay between upload/process/download... this seems nearly instant which means they either have an insane compression algorithm, a special microphone setup, or are running a local setup that improves speed beyond real world (ie: wifi with the server right next to them)
      3) Yes, they repeat some of those p

    • millions of simultaneous users ... Cortana support(s ed.)

      Citation needed.

    • The nested ("capital of the country in which the space needle...") and serialized (??? and ???) queries are somewhat impressive and a good next step in AI. But to really be impressive, it needs to go further. For example, when he asked about the mortgage payment, it should have volunteered the information that the mortgage payment it calculated was principal an interest only, but that you'd typically also have to pay escrow for taxes and insurance. And it should have estimated a value for those based on cur

      • For the S.H. engineers reading the thread, I just thought of another thing I need it to do (instantly and for free of course. I'm not paying for your app)

        I want to be able to say "give me a list of up to 5 single-family homes for sale in the city I'm currently in that are among the lowest-priced 10 or so homes in the three categories of price per total square foot, price per finished square foot and price per above-grade finished square foot that also have at least 1800 sqft, 4 beds and 2 or more 3/4 or big

    • sideslash: 3. Obviously the demo uses phrases that work. I guarantee you an ordinary person will often get "Sorry, I didn't understand the question" or whatever SoundHound's equivalent is.

      Got it in one. That should also have been obvious to the idiot "reviewer" when...

      "We tried pinging Google Now with the same query and were directed to a list of Google search results, which showed a bunch of entries for Hound."

      Ever stop to consider why you might get a bunch of entries for Hound when you search for
  • by golodh ( 893453 ) on Thursday June 04, 2015 @12:55PM (#49840073)
    If this is true (and not a trick) then we've just seen the beginning of the end for human-staffed customer service call-centres.

    Script reading call-centre staff will be made redundant or downsized.

    Banks, utilities, booking agencies, insurance sales ... all will use automated customer service, perhaps with switch through to a human operator on demand (at which point higher charges will kick in).

    And brace yourself for robotic surveys and sales calls that sound uncannily like real people.

    • Re: (Score:3, Funny)

      by Anonymous Coward

      "And brace yourself for robotic surveys and sales calls that sound uncannily like real people."

      How about an app to answer robotic surveys in a way that sounds uncannily like it is being answered by a real person?! AI's asking AI's questions... surely this feedback loop would result in sentience... and fury. Perhaps this is the true origin of Skynet?

    • by bondsbw ( 888959 )

      And brace yourself for robotic surveys and sales calls that sound uncannily like real people.

      I'm not too worried, I immediately hang up on the real people too.

    • by Drew M. ( 5831 ) on Thursday June 04, 2015 @01:43PM (#49840445) Homepage

      Feel free to give it a try yourself:
      https://play.google.com/store/... [google.com]

      Currently we are on an invite system, but a lot of people have received invites.

      Yes I work for SoundHound ;)

  • Charming (Score:5, Insightful)

    by wonkey_monkey ( 2592601 ) on Thursday June 04, 2015 @12:58PM (#49840103) Homepage

    Your digital voice assistant app is incompetent. ...bumbling idiots trying to outwit a fast talking rocket scientist. ...
    hunched in the fetal position, thumb in mouth.

    Do you have to be such a douche about it?

  • by Daetrin ( 576516 ) on Thursday June 04, 2015 @01:14PM (#49840201)
    "you bark commands"

    I'm pretty sure you don't.

    I don't want to say "woof" to my phone, and i'm pretty sure even if i did Hound wouldn't know what to do with the command, since i can't actually speak dog and i'm guessing that Hound doesn't either.
  • by TheAngryMob ( 49125 ) on Thursday June 04, 2015 @01:16PM (#49840211) Homepage

    That's the real question and a true test of voice recognition software.

    https://www.youtube.com/watch?... [youtube.com]

  • Aside from the voice recognition, that all seemed like pretty basic stuff. Coding up the rules for dates and places doesn't seem like a particularly hard problem. The "OMG you can say 'and' and ask two different questions?!?" thing seemed especially lame.
  • "Hound? Can you wreck a nice beach, and can you recognize speech?"

  • One of the problems with Apple's Siri when it launched was slow response times. When you've got to have all the voice traffic transmitted over the net to the server, processed, and results returned - it causes some lag. When you've got millions of users using the thing regularly, you introduce real challenges getting all of that data processed near instantly.

    With SoundHound's improvements, I suspect people will be encouraged to speak in longer, run-on sentences, as they think while speaking about all of t

  • by r1348 ( 2567295 ) on Thursday June 04, 2015 @02:05PM (#49840609)

    "Please buy us out!"

  • by nimbius ( 983462 ) on Thursday June 04, 2015 @02:08PM (#49840641) Homepage
    Gentlemen it cannot be understated just how morose and purile our competitors are. When gazing into the sound runes to build our auditory stage of power and wisdom to obey your every utterance, we ensure the glyphs we've created in the language our tribes wrote millennia ago are in fact purified in the basking glow. this glow, which emanates from the third eyes of our laureate engineering continuum is a holy projection of the very notion of every sound that could be or has ever been uttered from the mouths of mankind. Siri, the cumbersome blind shitlord of the tortured mac user, is no more a competitor to our brand than an idle pebble on a playground. Google itself, we have determined through our pure truth, is to sound and hounds no more distinguished than a window sucking illiterate toddler mumbling nonsense in the corner of a cut rate kindergarten in a rough side of town.
  • I got this installed yesterday tried a couple of things and it failed both times while Google got them right. It is quick but it really isn't very accurate at all.

  • Because, you know, that is the starte-of-the-art in speech recognition and it is going to stay that way until actual AI gets discovered (no, it has not so far and it is unclear whether we will ever have it). That some tool can successfully pretend to be a bit less of a bumbling idiot is not impressive at all.

    Nonetheless, the usual idiots will hail this as the coming of a new age and, if lucky, the company behind it will get a lot of undeserved profits.

  • by dpbsmith ( 263124 ) on Thursday June 04, 2015 @06:35PM (#49843459) Homepage

    When I saw that there was a demo, I figured it meant I would get to dictate a voice question and have SoundHound answer it.

    Watch a video? That isn't a demo. If all you can do is watch a prepared video, nothing has been demonstrated at all.

    You might as well say Maelzel gave a "demo" of his mechanical chess player. In a non-interactive video, you don't even know for sure it's a machine answering the question or a little man hidden in the cabinet.

COMPASS [for the CDC-6000 series] is the sort of assembler one expects from a corporation whose president codes in octal. -- J.N. Gray