Google Voice Speech Recognition Not Really Working | Hacker News Reader

Google Voice Speech Recognition Not Really Working

rashmash.com

12 pointsnader17 comments

You have a bit of an accent (no offense, so do I), and that would normally throw the voice recognition software off.

Yeah, listening closely you can see how it gets "55" from his "bye bye". The strong accent makes it quite hard for the software to guess the phonemes being spoken.

You can't have a "one size fits all" approach in voice recognition if there are people pronouncing exactly the same words in completely different ways.

Thinking about the United States and the variety of cultures / accents (mexican, chinese, etc.) I guess errors will happen very often.

This brings up an interesting problem with the future coding of Google Voice. How is it suppose to determine whether it's someone with an accent saying bye bye or someone without an accent saying 55. Even if the dielects are taken into consideration, that is still going to be quite a difficult problem!

Context. You could compute the log likelihood ratio of "bye bye" given the previous word versus the log likelihood ratio of "55" given the previous word. If that's not good enough you can generalize this to use the previous N words.

Google should have a large enough corpus of text at it's disposal to be able to approach the problem this way, and they very well may already be doing so.

I don’t see why you can’t have "one size fits all" recognition. It just requires more advanced profiling. The computer has to be smart enough to use all the words leading up to that one hard to recognize word. Not to mention other cues like grammar check (which would have revealed a lot of what went wrong in the example the author uses)

I’m not saying it’s easy. In fact, I’d be willing to bet money that I personally don’t have the programming chops to do it effectively. But I know enough to know it can be done.

Anecdote: the speech recognition works unsettlingly well for me.

I use my Google Voice number on my resume, to prevent recruiters from calling my cell phone directly. I've noticed that clear speakers with the classic "TV American" dialect are transcribed pretty well. However, if they have even the slightest bit of a different dialect, especially Indian or Southeastern US, or if they mumble a bit, then it's hilariously bad.

Stuff like this:

    hi  zack  this  is  anyhow  calling  from  a  car  
    and  i  actually  i  was  looking  part  of  the
    apartment  but  i  might  come  across  your  profile
    on  the  google  search  in  hi  i  understand  you're
    looking  for  approximately  400  caveman  income
    which  is  definitely  i  can  talk  to  you  but  i
    was  just  wondering  that  if  you  know  anyone  and
    some  yesterday  and  so  and  relatives  so  of
    because  the  anyway  i  was  looking  for  a  job
    for  their  business  hey  can  you  please  a
    deficit  to  me  on  my  email  address  is  and  and
    it's  shannon  N  T  S  as  in  sam  but  i  am  E  T
    oddity  of  I  T  dot  com  i  repeat  it  and  and
    at  the  car  dot  com  on  if  you  can  divert  my
    call  at  84  so  i'm  fine  i  need  to  little
    photo  a  photo  extension  8  if  i'm  not  i  repeat
    841  598  double  total  of  4  extension  8811  that
    would  be  great  thanks  so  much  for  your  time

My experience is that recruiters (and other people who leave a lot of voice mail) seem to be relatively easy for Google Voice to transcribe:

"matt hey this is todd johnson i work over dates 10 capital hey i wanted to see if you knew of anyone open for new projects we've got up i've done contract role that just popped in if you have a chance feel free to give me a call (206) 300-2120 to seattle based company python project 6 months see if you might be interested in that again this is todd johnson over it's 10 capital thanks"

Whereas casual callers talk faster and tend to get transcribed like this:

"hello it's sarah this is christal donna considering it's jamie at coming thanks bye"

What I really like is the feature in the new Android app that highlights each work of the transcript while playing the audio, and lets you skip to any word just by touching it.

Does anyone know whose engine they use? There are only a couple of Speaker Independent Speech Recognition Engines out there. Have they developed their own or licensed it from someone?

Google wrote their own engine after hiring away a bunch of the top guys from Nuance.

They started it with Goog-411, training it on requests like "Pizza in Oregon" where they'd know if they were right or wrong because people would request another listing if they weren't quite correct.

This is the next step of that, so far as I understand. They're giving it a broader base of phrases to work on. I imagine the ultimate goal is to throw it up against Youtube, so you can search for a phrase, and Google will give you YouTube videos where someone speaks it.

I heard the rumor once and lose the url. It says Google hires the original engineer who created Nuance Dragaon NaturallySpeaking. There is also a old rumor about one goal for Google's participation in WINE Project is to let Nuance Dragon NaturallySpeaking working on Unbuntu.

I think the reasonable assumption is that Google hires engineers and built a version of the engine on their own platform. So it may be a new engine in different architecture as the desktop PC version but with all the brains behind Dragon NatuallySpeaking.

As Voice matures and more people use it Google will continually get a larger dataset to draw information from. Perhaps it could feed into a better probabilistic transcription approach in the veins of Google Translate.

You are probably right but there's a certain enjoyable irony to Google getting dinged for this being they, more than anyone else, are responsible for devaluing the word "Beta"

Hopefully incidents like this will make them realize the word had an important use after all and that maybe it isn't the best idea to tag every product with it for a virtually indefinite period of time.

The technology being used to power my G1's "Search by Voice" feature is really surprisingly good. I would assume it's the same stuff. Maybe it was just the unusual names that threw it.

You mean everybody should have a name like "Bob", "Bill" ? :) I have a feeling that the Search by Voice Feature on the iPhone works better than on the website but that might just be by accident

Perhaps audio quality makes a difference? On the G1/iPhone it can sample audio at whatever rate it needs, rather than just having to take what comes over the line.

Reply on news.ycombinator.com