With the release of Pixel 2 Google introduced the Now Playing feature, which is able to recognize the music playing in the background. Now Playing constantly analyzes ambient sound, requires no Internet connection, the entire database of tracks stored on the device itself, the main feature of the new feature was the use of neural networks to identify tracks. The Google blog AI, the company announced that it has launched a similar neural network to the recognition Sound Search, located on the server side and accessible to everyone.
Now, ask the assistant “OK Google, what’s this song?”, you can get the same quick and accurate response, as in the case with the Now Playing:
For recognition in the Now Playing company uses a fingerprint recording of surrounding space, after which a similar fingerprint is compared to fingerprints in a database of the smartphone. In Now Playing is stored thousands of songs, the database is constantly updated, adding new tracks, while unpopular are deleted. This creates some limitations.
In the case of Sound Search database 1000 times more, allowing you to more accurately identify and track simply are more likely to withdraw at least some result.
Under the imprint means something like this:
We have a record length of 9 seconds obtained by using convolutional neural networks which is divided into 2-second chunks every 1 second. Meanwhile, the database stores the exact same prints. Now Playing with a two-phase algorithm compares the fingerprint to a database. In the first phase using the hybrid spatial partitioning and vector partitioning analyzes the entire database searching is more or less similar tracks. In the second stage among the candidates is more in-depth analysis of each two-second piece.
Due to too much base in the Sound Search the developers have improved the technology, is 4 times increasing the size of the neural network, the segments reduced to 1 second at intervals of 0.5 seconds. By the way, something similar to the voice recognition uses the company Yandex.
Share your opinion in the Telegram chat or in the comments below.