Google originally launched its patented cloud-based voice recognition system in 2007 to power its GOOG-411 information service. By gathering data from hundreds of phone calls, it improved the service and now uses it to power automated voicemail transcriptions for Google Voice as well as powering voice-based search for Android and iOS. The company is taking voice recognition one step further on Android by adding "personalized recognition" to its algorithm.
Google's current system is designed to work "regardless of gender, age, and accents, or variations in pitch, pace, and other factors." However, anyone that has used voice recognition software knows that training it to recognize your particular voice and speech patterns makes it far more accurate. In comparison, one look at any Google Voice transcript and you'll see Google's system isn't anywhere close to 100 percent accurate.
Google hopes to change that by adding optional personalized recognition to its Android Voice Search app. By opting in to the system, Google will store recordings of your voice search queries and voice commands in your Google account. These recordings will be used to build a recognition model for your particular voice and speech idiosyncrasies.
"Although subtle," wrote Googlers Amir Mané and Glen Shires, "accuracy improvements begin fairly quickly and will build over time."
The new personalized recognition feature is only available for Android 2.2 or higher—no word if the improvements will spread to other platforms. It's currently limited to US English, but support for new languages will be rolled out shortly. As an added bonus, the new version of Voice Search includes speed and performance improvements for name recognition, particularly over slower 3G or EDGE data connections.
Source: ars technica
