Last month, Google launched a feature that allows users to hum a tune to search for a song. This melody-style method of searching for music isn't completely new, since apps like Shazam allow users to look up songs they hear on the radio. However, Google is taking things a step further by letting users simply hum the song that they can't quite remember the words to.
The innovators at Google have been working on music recognition technology for years now. Google launched Now Playing on the Pixel 2 in 2017, a feature that brought low-power music recognition to mobile devices. A year later, the same technology was introduced to the SoundSearch feature in the Google app. Fast-forward to 2020, and Google has found a way to streamline searching for songs on the tips of a user's tongue.
Users can open the latest version of the Google app, tap the mic icon, and say "what's this song?" or click the "Search a song" button, and then start humming for 10-15 seconds. Google's machine learning algorithm will take the humming and try to match it to potential songs. The new feature is available through Google Assistant as well, by simply saying "Hey Google, what's this song?" and then humming the tune. The new technology is currently available in 20 different languages on Android, but only in English on iOS.
The Science Behind Google's 'Hum To Search'
Google says its Hum To Search technology works in contrast to other approaches to melody searching, which try to match a tune to a database of existing melody versions of the song, instead of finding the song directly. With Google's method, the technology "produces an embedding of a melody from a spectrogram of a song without generating an intermediate representation." Basically, this approach is supposed to be able to match a hummed melody directly to the original recording, without needing to match it to another hummed version. Google modified the music-recognition models used in previous technologies, Now Playing and SoundSearch, so that it can work with hummed recordings. That involved training the machine learning technology to produce embeddings for pairs of audio that contain the same melody, even if they are created using different instruments or singing voices. The result, Google says, is a system that reaches a high level of accuracy with a song database of over half a million.
This all adds up to the fact that Google has scientifically found a way to combat the issue of having songs that one can't quite remember fully. Granted, it may not be the most pressing issue of the 2020, but many will be familiar with the experience of trying to remember the name of a song and having it stuck in their head throughout the day, or longer. Even if a user can't remember the lyrics, or even the tune beyond a basic hum, Google now has a solution.
Source: Google