Voice recognition is becoming integral to the new mobile user experience, and to the new approaches to search, as enshrined in Apple Siri and Google Now. Microsoft is playing catch-up, but claims to have doubled the speed at which speech is recognized and results returned.
The technique, which is now being rolled out from the R&D stage to commercial web data centers, also claims to improve the accuracy of recognition by 15%, by using a computational model which mimics the way the human brain works.
Microsoft’s vast R&D unit has turned its brains to applying ‘deep neural networking’ to speech recognition. According to CNet, Michael Tjalve, a senior program manager in Microsoft’s speech technology program, claimed: ‘For a normal sentence, you will have one less word to correct.’ The speed of getting voice activated results from the Bing engine will double; speech translation to text will be almost instantaneous; and word error rates will fall from 16% to 13.5%.
However, the researchers did not offer direct comparison with Apple’s and Google’s technologies, saying they used different methods and measurements. To get its own improvements, Microsoft
replaced the acoustic model in its speech recognition technology, making it less susceptible to mistakes when speech patterns are variable than the currently used method the ‘Gaussian mixture model’. The new approach relies on the new levels of computational capacity which are becoming available to drive speech and artificial intelligence engines, which in turn are driving new forms of search.
Microsoft began rolling out the update to data centers in the US in April, and plans to have completed the process within the next few weeks, but did not give plans for international roll-out.