Using Bass-line Features for Content-Based MIR

We propose new audio features that can be extracted from bass lines. Most previous studies on content-based music information retrieval (MIR) used low-level features such as the mel-frequency cepstral coefficients and spectral centroid. Musical similarity based on these features works well to some extent but has a limit to capture fine musical characteristics. Because bass lines play important roles in both harmonic and rhythmic aspects and have a different style for each music genre, our bass-line features are expected to improve the similarity measure and classification accuracy. Furthermore, it is possible to achieve a similarity measure that enhances the bass-line characteristics by weighting the bass-line and other features. Results for applying our features to automatic genre classification and music collection visualization showed that our features improved genre classification accuracy and did achieve a similarity measure that enhances bass-line characteristics.