Mood music

Listeners will no longer have to search for the song to fit the moment if Glasgow researchers succeed in classifying its emotional content.

While most MP3 players enable you to carry around your entire music library of thousands of songs, sometimes you only want one perfect playlist.

Researchers from Glasgow Caledonian University have tuned in to this need and developed software that can classify the emotional content of contemporary music.

‘The main aim of this project is to develop reliable signal processing algorithms that can automatically extract from digital recordings of music the emotion that it elicits in the listener,’ said Dr Don Knox, a signal processing and audio engineer at Glasgow Caledonian and the principal investigator on the project.

‘Everyone nowadays has a mobile music player and like everyone else, with the sizing of songs on it, I generally find myself wondering, “What do I want to listen to today?” The only options open to us at the moment are textual information in tag fields that describes the music in terms of genre, artist or track type.

‘We’re looking at ways of accessing the music in these large collections so that you can listen to what you’re in the mood for.’

Spectral frequency

The researchers will extract and analyse features such as pitch, tonality and musical structures, known as the spectral frequency in signal processing, to classify music by emotion. Existing methods have only been tested on classical music.

‘It’s a very new area. There is only one study that I’m aware of that attempted to do this and what they did is extract fairly basic features from the audio file, such as spectral intensity, tempo and rhythm,’ said Knox.

‘From the point of view of music composition and performance, composers understand how to manipulate mode, tonality and key to get a certain effect in the listener. These features of musical structure are not extracted from the audio in any other approach, but this is what we want to do.’

Knox will work with Prof Raymond MacDonald, Glasgow Caledonian’s music psychologist, to carry out tests.

‘The project will require that we measure the emotional effect of a collection of music on a panel of at least 20 listeners, using means that are established in psychology. It can be in the form of verbal or pictorial self-report from the listeners at the same time as taking some physiological measurements, such as galvanic skin response (monitoring the change in the electrical properties of the skin in response to stress) and heart rate,’ said Knox.

The engineers will extract ‘detailed’ features from the frequency spectrum, analyse them in relation to the information gathered from the psychological tests and look for a correlation between the two. Knox said it remains to be seen how well they will be able to map the extracted signals to existing models of emotion in the music psychology field.

‘What they do is map levels of something like stress and arousal, and particular emotions will appear at specific points on those two axes,’ he said.

Changing mood

‘We want to use the measures of musical structure to make the process more accurate, but one of the challenges is that the mood may change during a song.

‘However, physiological measurements are useful for pinpointing the moment in time at which emotional response changes.

‘If we can get that information and study it against the music signals, we can hopefully have a map of how the emotion changes across the duration of the song.’

Knox expects challenges. For instance, the project is not trying to extract the meaning of lyrics, which means that a song with an exuberant melody but depressing lyrics could be wrongly classified. He also expects to make some interesting findings.

‘This research will quantify how big a role lyrics have as opposed to the music, and also reveal something about the subjectivity of emotion,’ said Knox.

‘Previous studies in music psychology suggest there is a surprisingly large amount of agreement about what the emotion of the music is.’

As well as creating software that would allow listeners to choose music based on emotion, and possibly recommend music on this basis, Knox said the research could also help music-makers.

‘We could turn it on its head and work with composers who compose music for a certain emotional effect,’ he said. ‘If we can extract features from the piece of music, then we can feed that back to the composer and tell them how effective he or she has been.’