Originally, my PhD focus was on the automatic recognition of chords from musical audio. However, my thesis is now likely to be extending some of this work to higher-level Music Information Retrieval (MIR) tasks. A brief (non-technical) description of my current work is below.
A lot of work has been done in chord recognition in recent times, and for the first year or so of study I was working on building a cutting-edge system for chordal analysis. It took a while to get up to speed with the work, but I think we’re there now!
With so many great systems out there already, what makes ours different? We have used machine learning techniques to ensure that our recognition algorithm as versatile as possible, by learning from various sources. One such source is online chord databases, where millions of partial chord annotations can be used to learn about diverse genres for which there is currently no good data.
Another concern when dealing with data of this size is scalability. We’ve made sure the system we’ve been developing can be run on a home PC in a reasonable amount of time for practical use. So in answer to the above question, scalability and flexibility are the core of our innovations. We’re currently working on a version to be released – watch this space!
What use is a chord recognition system? Well, it’s quite nice by itself if you’re a musician to help you play songs, but chords can be used for other things too.
One thing we’ve looked at recently is using chords (amongst other features) to predict how well songs will do commercially before their release. This resulted in a workshop paper at MML 2011 (see publication page for the paper) but also a website documenting our performance and interpreting the results – www.scoreahit.com. Head over and take a look!
Mood is another thing surely affected by chords. If a song contains only minor chords it has a more melancholy sound than one using sorely major chords. Do diminished chords imply an anxious feel? We’ve been investigating some of these questions and others relating to mood prediction from audio. Unfortunately I don’t have a flashy website for this work yet, but there is a paper in the publications section where you can find out more.
This work is supported by the Bristol Centre for Complexity Sciences.