Google may have acquired Geoffrey Hinton’s DNNresearch and is now using his technologies to power its Google+ photo search features, but the academic work Hinton did on deep neural networks (DNN) is now also helping Microsoft to improve its speech-recognition systems. Microsoft today announced that it is using DNNs to double the speed of its speech recognition engine for Windows Phone while bringing down its word-error rate by 15 percent. Bing Voice Search, the company says, now also works far better in noisy conditions.
For now, these improvements are only available for users in the U.S.
Microsoft says it quietly started rolling this new system out to Windows Phone users over the last few weeks. The new system is the result of the Bing Voice team working closely with Microsoft Research, the company’s network of 13 research labs that work on anything from improving cell phone battery life and machine learning to research in game theory and economics.
DNNs, Microsoft says, help researchers build a smarter acoustic model to represent the acoustic representations of a language. Essentially, the idea is to build a model of how the brain listens to and interprets speech. You can find more info about how Microsoft uses DNN here.
There can be little doubt that voice recognition is a pretty hot area right now. Google, with its conversational search feature, is currently leading the way, but Apple (with Siri), Microsoft and a number of startups like Maluuba are also all working on products that use voice recognition, natural language processing and other techniques to get users just a little bit closer to the “Star Trek computer” ideal.