Turns out we’ve all been working on the wrong voice recognition technology for the past 30 years. Google thinks that voice search is what’s required and has filed for a patent, number 7,027,987 which describes how it thinks a search engine should be used when voice is the only input, like on a mobile phone (actually SMS input to search engines might be better).
Interestingly this was filed back in February 2001 and has only just been granted, so this is not so much what Google is thinking now, but what Google was thinking back then.
Most digital media news agencies ran the piece suggesting that this is all about being able to speak searches into our phones, and that’s pretty much what the technique is all about. It focuses on just what type of strategy the remote server should adopt in dealing with a search which it hears in an ambiguous way. It attaches a probability to each of the words and word combinations that it thinks it heard and then searches the most ‘probable’ thing you said. It can work off different probability databases and gave the example of English spoken in a German accent getting different probabilities from English spoken in a French accent.
It would be particularly frustrating for a mobile user to speak into the phone asking for local restaurants only to be told after a long wait and a savage addition to his or her phone bill, the answer to something entirely different.
The truth about voice recognition is pretty well established and owned by IBM. It has operated some form of voice lab for 30 years and more or less any patent that is applied for in this technology you will find included in IBM’s patent pool. Voice recognition is easier to do on a voice that the software has been trained to understand, and it is even easier on limited vocabulary, neither of which helps a search engine. Back in 1981 IBM had a virtually perfect system for voice recognition, but the only problem was that the processor that you needed cost over $1 million back then.
Times have changed and we should expect processors and memory to be cheap enough to put voice recognition into internet held servers, which are attached to mobiles or other portable devices. It is thought that Sony is working on a real time PlayStation Portable voice service that will translate spoken language on the fly, from one PSP to another remote player that is sharing a networked game, using the huge power of the PS3 Cell chip to handle the translation. That would be more interesting than voice search we think, which is almost certainly doomed to failure. Existing Google searches often give thousands of hits, and the first few are usually useless, so how it could hope to get to a single, transferable hit using voice, we don’t know.