How Machine Learning is Revolutionizing Our Music Experience
Music is a major part of my identity, and several people across the world will say the same, speaking to its impact on their lives. Songs manifest the ability to cultivate our curiosity, keep us energized throughout the day, and provide an escape from reality. We all seek to have this type of experience, and Spotify is giving it to us through their incredible personalization features. The most notable one, in my opinion, is their discovery weekly playlist. Every Monday, 70 million Spotify users can look forward to enjoying a playlist of 30 songs that match their music taste. As an electro-indie fan, I have truly “discovered” from the Discover Weekly playlist my favorite songs. I can definitely say that Spotify knows my music better than anyone else, but how? How is Spotify able to capture our own unique preferences? The answer: an underlying Machine Learning algorithm that can explore the vast realm of music to deliver us the music we want to hear.
Here is an overview of a vital algorithm that Spotify has implemented to make this feature possible:
Convolutional Neural Network (CNN): A CNN can be pretty complex to wrap your mind around. How does it exactly work? Every machine learning model needs input, and for Spotify’s CNN it is an array that contains information about the frequency, duration, and amplitude of a particular note during a song. After organizing the important components of the song in a matrix/array, it is multiplied by another array called a filter to determine the presence or absence of a certain feature of a song (tone, melody, pitch, mood etc.). For instance, if the product of heavy bass filter array and the song input array is zero, then we can assume that this is probably not a Bassnectar tune. After this iterative process, a global temporal pooling layer is applied to generate insightful statistics like the mean and max occurrences of a certain aspect of a song such as a chord progression. This only helps to further categorize a genre of the song from the audio signals, and is further used as input into what is called a fully connected layer. This computes a single dimensional array consisting of values that specifically correspond to that song. Now that our morning “pump ups”, chill rewinds, and summer jams can all be represented by an array, Spotify can find 30 other arrays AKA songs that have the closest matching values. This what allows Spotify to keep fueling our musical passion and expose us to sounds we might not ever have encountered. I am fascinated to see how Spotify will leverage neural networks and Machine Learning to further enhance the way we connect with music.
Read more from RebellionResearch.com:
Written by Jason Scanlon & Edited by Alexander Fleiss
We just sent you an email. Please click the link in the email to confirm your subscription!