Hum to Search is a machine-learned system within Google Search that identifies songs from hummed or sung melodies.
Hum to Search is a feature launched by Google that allows users to find a song by humming or singing its melody. Unlike previous methods that relied on databases of pre-existing melody-only or hummed versions, this system directly matches a hummed melody to original polyphonic studio recordings. It achieves this by producing an embedding of a melody from a spectrogram without generating an intermediate representation, simplifying the database and allowing for constant updates with new releases. The system leverages machine learning models, initially adapted from Google's Now Playing and Sound Search technologies. It uses a neural network trained with pairs of hummed/sung audio and recorded audio to generate embeddings where similar melodies are close in embedding space, regardless of instrumentation or voice. Training data was augmented with simulated hummed melodies generated using SPICE, a pitch extraction model, to improve robustness for humming and whistling. The model's accuracy was further enhanced by refining the loss function to account for model confidence across various training examples. This technology is integrated into the Google app, enabling users to identify songs by tapping the mic icon and saying "what's this song?" or selecting the "Search a song" button. The current system boasts high accuracy on a database of over half a million songs, which is continuously updated.
Integrations: Google Search, Google App
Platforms: Web, iOS, Android
View full Hum to Search profile on Tools-Radar | Browse Voice & Audio tools | Alternatives to Hum to Search
Tools-Radar is a free directory of 10,000+ AI tools — discover, compare, and choose the right AI software for your needs. Visit tools-radar.com