Automated Mood Quantification 1 Running head: AUTOMATED MOOD QUANTIFICATION Automated Mood Quantification of Contemporary Music

This study investigated the possibility of automatically quantifying mood according to two parameters, energy and tension, with values similar to those that would be assigned by humans. Humans completed an Activation-Deactivation Adjective Check List on fifty 20-second music clips with moods varying across samples but uniform within samples. Human measurement of energy and tension were calculated from the survey results. Thirty features relevant to mood were extracted from each of the clips using the Marsyas framework. The features were used as input to neural networks in a variety of configurations using the extracted features paired with the centroid of the human measurements to create training data. Most configurations performed slightly worse than human performance in the median, and with significantly higher dispersion. One configuration had better median performance than human evaluation with acceptable dispersion between the first and third quartile, but a large dispersion across the entire set of results. While these initial results are promising, further refinement will be needed to apply the techniques to real-world systems. © 2006 Chris Cooke, All rights reserved. Automated Mood Quantification 3 AUTOMATED MOOD QUANTIFICATION OF CONTEMPORARY MUSIC Introduction When someone sits down to listen to music, how does he decide what to play? In some cases one may have a particular musician or genre in mind, but sometimes one wants to hear music of a certain mood. Since mood varies within a genre or the work of a particular musician, play list selection in this case can be problematic unless the listener has extensive knowledge of the available library holdings. Consider the case of an online music site with hundreds of independent artists and thousands of songs, with more of each being added daily. The ability to build a play list based on mood under such conditions seems hopeless without automation. Music information retrieval based on metadata such as artist name and song title has become routine. The chief problem is associating the metadata with a track in the first place, but with the advent of ID tags and databases such as the Gracenote database (Gracenote, 2006), this issue is solved even for the music of serious amateurs. Less work has been done in extracting information from the audio content of the song itself. Research into automatic genre classification shows promising results, but little as been done to detect mood. Feng et al. (2003) used tempo features to classify mood into four basic categories. Liu et al. (2003) used a more sophisticated set of audio features to classify acoustic music clips into one of four moods derived from Thayer’s model for mood and arousal. Thayer (1989) proposed a model for mood that consists, in its basic form, of two dimensions of arousal, energy and tension. Liu’s team quantized a planar segment of these dimensions into four distinct moods by pairing high vs. low energy with high vs. low tension (Liu’s paper uses the term “stress”.) Liu’s team had some success automatically classifying 20-second sound clips into one of these four © 2006 Chris Cooke, All rights reserved. Automated Mood Quantification 4 categories using spectral and time-based features as input to a hierarchical Gaussian Mixture Model, which used probabilistic methods to build a decision tree for classification. While Liu’s work represents a good first step, one might ask if a four-bin classification is too broad for practical play list construction based on mood. Rather than classifying a song into one of four moods, there might be some value in describing the degree of energy and tension arousal with some metric. Such a measure could then be associated with the audio file and be used to match a similar metric derived from the listener’s mood, perhaps using a distance algorithm or a fuzzy expert system. The central question, then is given a set of extracted features, can a machine-learning algorithm create a measure of energy and tension arousal that is comparable to that of a human listening to the same clip?