Evaluating Algorithmic Prediction of Emotion in Music

Algorithms play an increasingly important role in our interactions with digital services. Companies like Spotify and Pandora use algorithms to identify mood-relevant music from massive databases of over 100 million tracks and deliver them to users through personalized playlists. To achieve this, they rely on software tools to analyze relevant musical properties—like timing (slow–fast), intensity (soft–loud) and pitch (low–high). However, validation studies of common music analysis tools reveal inaccuracies in their performance—raising question to their value for emotion-based music recommendation. To improve the quality of mood-based personalization, this project will evaluate the reliability of music analysis tools with statistical modeling—clarifying their accuracy at predicting participants’ perceived emotions from listening experiments. The project outcomes will provide an important step toward improving algorithms’ predictions of emotion in music, along with users’ experiences with online music services.

Faculty Supervisor:

Michael Schutz

Student:

Partner:

Durham University

Discipline:

Sociology

Sector:

Education

University:

McMaster University

Program: