Google has introduced a new version of ‘Sound Search’, its server-side music recognition service that works alongside Google Assistant. According to a post on the company’s AI blog, ‘Sound Search’ is now powered by the same deep learning technology that’s used by Google’s ‘Now Playing’ offline music recognition service.
For the unaware, ‘Now Playing’ was introduced last year, at the launch of the company’s Pixel 2 and Pixel 2 XL smartphones. It works by sampling a small part of a track already in the music database, and thus, can run even without a network connection.
As per Google, ‘Now Playing’ not only works great as an on-device music recognizer, but it was found that it exceeds the accuracy and efficiency of the previous version of ‘Sound Search’ as well.
The new ‘Sound Search’ can be used directly from Google Search app or Google Assistant. Users have to simply start a query by asking “Hey Google, what’s this song?” (provided a track is playing nearby), and the service will return the name of the song that’s currently being played.
Google further explains that the core music identification process works by “generating an eight-second long ‘fingerprint’ of the track’s features, which is then projected into a sequence of low-dimensional embedding spaces having seven two-second clips separated by an interval of one second.”
Watch: Google Pixel 2 XL First Look
This is followed by a two-phase matching process, with the first involving the usage of a fast (but inaccurate) algorithm, and the second based on detailed analysis. Google says that there’s still room for improvement in the system, as it sometimes fails to recognize tracks in noisy environments. There are plans to make the system even faster, and improvements will continue to come in future.