Google announced an update bringing neural network integration to Sound Search. This has been adapted from the same technology powering the Now Playing feature of the Pixel 2 phones. Available from either Google Search or Google Assistant, Sound Search is a feature that allows users to detect and identify songs. From now on, Sound Search will use modified version of Now Playing's neural network algorithms to identify songs, improving both the accuracy and speed of song matches.
Now Playing is a functionality that originally launched alongside Google’s Pixel 2 handsets last year. Described as a portable version of Sound Search, it is a feature that allows a compatible phone to automatically identify songs even while operating in sleep mode. Instead of relying on an online database like Sound Search does, Now Playing works by matching songs to an offline database stored directly on the device. When the phone detects a song playing, it records an eight-second clip which it uses to search its on-device database for a match. Because the recorded audio is often distorted with background noise. Google employs a two-step matching process to identify songs. The first step consists of a quick search to narrow down the results to a handful of songs. Next, the phone will directly match the recording with the selected songs to determine the best match.
Compared to a traditional Sound Search through the Google App, Now Playing’s methodology benefits from low energy usage and the ability to operate without constant internet access. The downsides include a smaller library of supported songs, with Sound Search being able to identify about a thousand times as many songs as Now Playing can. In order to integrate this search mechanism into Sound Search, Google made a few tweaks. To compensate for Sound Search's larger database, Google not only increased the sampling rates but also the size of the neural network powering the entire affair. Because of this, Google was also able to lower the amount of time needed to identify a song. Alphabet's subsidiary also adjusted the algorithm to account for song popularity, meaning that Song Search will be more likely to identify popular songs even with poor audio samples.