Have you ever had a tune stuck in your head but couldn’t remember the song’s name or lyrics? Google Search has a clever solution for this: humming to search. But how does Google actually identify a song just from a hum? It’s all thanks to sophisticated machine learning models that understand the unique “fingerprint” of a melody.
Imagine a song’s melody as its fingerprint – distinct and recognizable. Google has developed machine learning models capable of matching your hum, whistle, or singing to the correct musical “fingerprint.” When you hum a tune into Google Search, these models spring into action, transforming the audio into a sequence of numbers. This number-based sequence acts as a unique identifier for the melody.
These models are trained extensively using diverse audio sources, ranging from people humming, whistling, or singing, to studio recordings of songs. Interestingly, the algorithms are designed to filter out extraneous details like accompanying instruments, vocal timbre, and tone. By stripping away these elements, the models focus solely on the core melody, represented by the number sequence – the song’s essential fingerprint.
The magic happens when Google compares this melody fingerprint against a vast database of thousands of songs from across the globe. This comparison occurs in real-time, identifying potential matches almost instantaneously. Think about the song “Dance Monkey” by Tones and I. You can recognize it whether it’s sung, whistled, or played in its original studio version. Similarly, Google’s machine learning models can recognize the underlying melody from a studio recording and match it to a hummed audio input.
This innovative feature builds upon Google Research’s prior work in music recognition technology. The journey began with “Now Playing” on Pixel 2 in 2017, which utilized deep neural networks for low-power music recognition directly on mobile devices. In 2018, this technology was integrated into the SoundSearch feature within the Google app, expanding its reach to a catalog of millions of songs. “Hum to search” takes this technology a significant leap forward, enabling song recognition even without lyrics or the original song – all it needs is your hum.
