Users of Google's Voice Search Android application can now "train" the software to understand their own voices.
Voice Search lets users speak search terms into a device instead of having to type them. In Google's cloud infrastructure, the spoken words are translated into text and run through the Google search engine. Since Voice Search was introduced in 2008, Google has addressed differences among voices by making its speech models broad enough to understand different accents, ages and genders.
With the latest version of the app, available now for devices with Android versions 2.2 and higher, users can make the software analyse all their verbal requests and learn the idiosyncrasies of their voices. Google Voice will gradually build a speech model for the individual, making it more capable of foolproof, instant understanding, Google said.
"Although subtle, accuracy improvements begin fairly quickly and will build over time," said a post on Google's official blog by Product Manager Amir Mane and Member of Technical Staff Glen Shires. The post includes a barcode that, when scanned, can direct an Android device to the Android Market page where the new version is available.
Users of the updated app will be asked if they want to opt into the personalisation system. If they do, recordings of the user's search requests will be associated with his or her Google Account. The feature can be enabled or disabled at any time, and the voice recordings can be disassociated from the account.
Google already records and stores spoken requests from Voice Search for as long as two years, but the recordings are not linked to the speakers' names or Google Accounts, spokeswoman Nadja Blagojevic said via email.
So far, Google has been able to satisfy most Voice Search users without providing the extra fine-tuning of personalisation, Opus Research analyst Dan Miller said. The company has been improving the accuracy of the speech recognition gradually using a broad range of voice samples. Earlier this year, Google said its search speech recognition was about 70 percent accurate and that one in four searches on Android 2.0 was done via Voice Search.
"It's a very complex task to recognise voice accurately, and you get to a point of diminishing returns, given the current model," Miller said.
It's likely that only a small number of people will use the new feature, at least at first, but it is a significant advance, according to Miller. It also will help to set Voice Search apart from rival services, he wrote in a blog post on Tuesday.
"It marks the recognition by Google that ... we really aspire to the Star Trek way of people interacting with machines," in which a person can fully control a computer by talking to it in natural language, with no misunderstandings, Miller said. "That really is going to rely on a certain level of personalisation."
In terms of security, stored recordings of search requests -- even if associated with the speaker's Google Account -- probably wouldn't be sufficient to allow criminals to fool voice-based user authentication systems, Miller said.
The personalization feature is available in English in the US now, and Google said it plans to expand the offering to other countries and other languages in the near future.