Google copy Siri (probably)

[edutech4schools]

Google is readying a Siri rival called Majel - The Inquirer

 

From the artical

I like to think of it like Stephen Hawking because it is extremely smart and you can interact with it naturally, but it is incapable of physically doing much.

[SYNACK]

Grrr, macolittes think Apple invent everything, voice rec was around long before Apple got involved by 'buying' another companies idea and shutting everyone else out of it. Other phones were able to pull this off on the device, even the idea of cloud offloading was used before on the Windows Phone platform. The thing that was different and created by the previous owners of Siri were the more integrated approach with natural language processing - it still sucks, I have tried it and it is about as contextually intelligent as your average four year old. Still, this 'leap' of inovation is really just a rubbish implementation of wolfram alpha, which is where is gets most of its intelligence from.

 

I like the fact they link it back to the startrek computer systems though as it is much more honest given that it is practically the same thing, a finite state machine with a NLA front end. I also imagine google will do a much better job of it being able to tap into their language interpritation stuff as well along with Wolfram Alphas.

[glennda]

I read somewhere proper siri was coming to android soon...

[edutech4schools]

I do think natural language is a way off, not used siri so cant comment on that, but must say my android and xbox kinect get my voice commands 99.9999% of the time. Having my android connected to my blue tooth car radio and its only miss dialled 1 number wrong in over a year, which is fantastic.

[mac_shinobi]

Grrr, macolittes think Apple invent everything, voice rec was around long before Apple got involved by 'buying' another companies idea and shutting everyone else out of it. Other phones were able to pull this off on the device, even the idea of cloud offloading was used before on the Windows Phone platform. The thing that was different and created by the previous owners of Siri were the more integrated approach with natural language processing - it still sucks, I have tried it and it is about as contextually intelligent as your average four year old. Still, this 'leap' of inovation is really just a rubbish implementation of wolfram alpha, which is where is gets most of its intelligence from.

 

I like the fact they link it back to the startrek computer systems though as it is much more honest given that it is practically the same thing, a finite state machine with a NLA front end. I also imagine google will do a much better job of it being able to tap into their language interpritation stuff as well along with Wolfram Alphas.

 

I remember 'Dragon naturally speaking' for the pc, tried a 30 day trial of it and you had to train it and even after hours upon hours of training it still sucked !!

[SYNACK]

I remember 'Dragon naturally speaking' for the pc, tried a 30 day trial of it and you had to train it and even after hours upon hours of training it still sucked !!

 

Yes, that was because it was trying to be 100% accurate with no context and full grammer/structure support, it was also at a time when you were limited in the number of voices you could train with. You can make a hell of a lot more assumptions when dealing with simple question/answer stuff and the methods have changed. Things like Google Voice can run their learning algorythms over millions of different voices and so real technology companies like Google and Microsoft can chuck the whole lot into an expert system to learn accents and tricks from a totally different size dataset. MSR (Microsoft Research - I find it funny how close the acronym is to USR) for instance has spent lots of time messing with this kind of stuff which has found its way into all sorts of products, Office, Windows, Exchange, Link (or whatever they call unified comms server now) and more importantly xbox games and windows phone. Several of these are datacentre backed and play the same tricks that Google does with huge ammounts of data.

 

I'm not saying that there is not room for other providers but these two have had a huge head start with the data, and the data is in wide use cases, phone messages, voice commands, data entry. Apple does have plenty of iFollowers but they have plenty of catching up to do and in contrast most of their data will be soundbites rather than longwinded monologues or conversations.

 

Stuff like Dragon have also started going the centralised route at least with their mobile apps, it all comes down to who had the most and best data along with the smartest and most unconstrained people. This is what Google voice was about, anyone who thinks they offered it free out of the goodness of their hearts is delusional, its about the data, who called, when they called and what they said along with how they said it.

 

Some problems like voice rec are just that complex that you need billions of samples in your dataset to crunch the numbers in order to solve. Its only now that this kind of analisys is computationally possible and the way that people are thinking about them (in datamining terms) that this kind of stuff becomes really solvable. The interesting bit is what else all this data could be used to solve with the right algorythm and enough CPUs.

[dhicks]

Some problems like voice rec are just that complex that you need billions of samples in your dataset to crunch the numbers in order to solve.

 

I was interested to see the way Gogole's spell checker works - it's pretty much a one-page Python program that reads in a bunch of standard texts (the top 100 most popular from the Gutenburg Project), which it assumes are spelt correctly, then compares each word in a document to that set of words.

 

I wonder if Google are doing pretty much exactly the same for voice - get a bunch of standard texts which you have both text and audio for (I'm guessing the top 100 from the Gutenburg Project would do there, too), assume they are pronounced correctly, and then comapre any incoming audio to that. My guess is that to make sure you are recognised by modern speech recognition software you're best off speaking as though you are in a Austen or Dickens novell...