SIAM, 11 / 10 / 2010 gene kogan

Research: Given a song, find the point on the valence-arousal plane which best represents it. Development: Given a point on the valencearousal plane, deliver music which will induce that mood.

Song is described by a vector of features

Machine “discovers” valence/arousal label
Learn by example

Constraints

Song

• Input • Extract features describing the song • Predict valence and arousal • Evaluate error

Features

Classifier

Label

Inherent subjectivity of emotion

Data collection
Felt vs. perceived

Signal-based vs. “secondhand” vs. metadata

Psychoacoustics in 60 seconds

Non-tonal: MFCC, sones, etc

Tonal: Chroma, tonality

Model using support vector regression

Evaluate performance using cross-validation

High-level, secondhand features Temporailty

The trivial way

Clustering

Modeling residual/deviation