The Last.fm headquarters in London felt a bit like a university lab for a few hours as Tom Walters visited this Tuesday. We persuaded him to tell us a bit about his research, and he explained and demo’d his pre-Google work on auditory perception, parts of which he is now applying at Google. In particular I was excited to learn that the source code of his auditory model is actually available at SoundSoftware… that seems to beg being cast into a Vamp plugin! But not only Mark Levy and I were interested in his stuff, it seemed as if the more techy part of the Last.fm crew, too, were all ears. Thanks Tom, a very pleasant afternoon.

Both McGill and New York University have announced the release of separate new chord ground truth data sets. Taemin Cho of New York University announced around 300 songs, among which all the RWC pop song collection. The data is going to live partly on the existing RWC website. Ashley Burgoyne presented a set of 1000 songs from the billboard charts, 500 of which are published immediately, and 500 will follow in staggered chunks over the next few years in order to allow for unbiased testing. The data is available on a dedicated website, and what’s coolest is that they provide both Echonest and my very own NNLS Chroma features along with the symbolic chord data, so that people can immediately go do stuff with them, even if they donR…

Today at ISMIR Masataka Goto finally presented Songle.jp - our web service for “active music listening and content-based music browsing”. It’s awesome. I’m really proud I’m part of it: I did the chord, beat and bar annotation code. But, of course, there’s much more to it: it involved not only the other music features (segmentation melody detection), but also the whole infrastructure and web programming. The guys in Japan have really been working towards this day for two years - it’s a very ambitious product. It’s really fun also, and you can only understand if you actually check some songs out… why not start with this.
Or, if you’re more a theoretical kind of guy, th…

In the heat of the moment I’ve got two things wrong in my paper entitled Timbre and Melody Features for the Recognition of Vocal Activity and Instrumental Solos in Polyphonic Music. First of all I apologise to Ferdinand Fuhrmann, Perfecto Herrera and Xavier Serra for not noticing their paper Detecting Solo Phrases in Music Using Spectral and Pitch-related Descriptors. Journal of New Music Research, 38(4), 343-356. Hence, contrary to what we state in our paper, we are not the first ones to have tackled solo activity detection. We might still be the first ones to have done it on popular music data, and I hope that we have added some valid contributions to the task. Secondly, there is two sentences from the original submission that…

As a supplement to my ISMIR paper, I now opened up the Structural Change code in a Last.fm github repository. Thanks to Last.fm for letting me do this. As anyone reading the paper may have spotted, it’s a relatively simple concept, and also similar to previous work by Sapp, Streicher and Foote. But I believe that using the change vectors at different time scales as feature vectors could result in genuinely new uses for existing audio features. My colleague Marcus has helped me make the code much more beautiful than it was only a week ago (thanks). It consists of two C++ (.hpp) files (one of which you don’t need if you’re not working with Vamp plugins). But let’s assume that you actually have a Vamp plugin with a vec…

I was invited by Dan Tidhar to present at last week’s visualisation workshop at King’s. It’s been a very pleasant experience, and and we discussed quite a broad range of topics, triggered by the different presentations. Mine was possibly the least interesting one, I just gave an overview over what kind of visualisations my work has produced over the last few years. Links: my slides, Audio Flowers, NNLS Chroma, Song Prompter. We also got a nice overview from Elaine Chew about her work of the last decade or so, King’s resident artist Michael Takeo Magruder showed us his techy visualisation techniques, in which he tries to keep the data “pure”. The most junior presenter was Mats Küssner. In what was maybe the most interesting presentation (from my point of view) Mats reported his results of a experiment in which people were asked to draw pitch and loudness of a sound on a pad. Especially the differing behaviours of musicians vs. non-musicians came quite unexpected. Beautiful research. There were other attendees from the Natural History Museum, and University College, hope I can confirm the names soon. In general, well done to Dan Tidhar and Daniel Leech Wilkinson for organising this, I really enjoyed it!…

Abstract: We propose the novel audio feature structural change for the analysis and visualisation of recorded music, and argue that it is related to a particular notion of musical complexity. Structural change is a meta feature that can be calculated from an arbitrary frame-wise basis feature, with each element in the structural change feature vector representing the change of the basis feature at a different time scale. We describe an efficient implementation of the feature and discuss its properties based on three basis features pertaining to harmony, rhythm and timbre. We present a novel flower-like visualisation that allows us to illustrate the overall structural change characteristics of a piece of audio in a compact wa…

Abstract: This paper describes a public web service for active music listening, Songle, that enriches music listening experiences by using music-understanding technologies based on signal processing. Although various research-level interfaces and technologies have been developed, it has not been easy to get people to use them in everyday life. Songle serves as a showcase to demonstrate how people can benefit from music-understanding technologies by enabling people to experience active music listening interfaces on the web. Songle facilitates deeper understanding of music by visualizing music scene descriptions estimated automatically, such as music structure, hierarchical beat structure, melody line, and chor…

Abstract: We propose the task of detecting instrumental solos in polyphonic music recordings, and the usage of a set of four audio features for vocal and instrumental activity detection. Three of the features are based on the prior extraction of the predominant melody line, and have not been used in the context of vocal/instrumental activity detection. Using a support vector machine hidden Markov model we conduct 14 experiments to validate several combinations of our proposed fea- tures. Our results clearly demonstrate the benefit of combining the features: the best performance was always achieved by combining all four features. The top accuracy for vocal activity detection is 87.2%. The more difficult task of detecting ins…

I’ve been awarded a 5 year research fellowship from the Royal Academy of Engineering for my project “Software Systems for Computer-Aided Music Understanding”. My work will be carried out at the Centre for Digital Music, part of Queen Mary’s qMedia centre and the School of Electronic Engineering and Computer Science. No decision has been made yet as to when I will start the work [edit: it's going to be January 2012]. We’re currently doing really interesting work in Last.fm’s MIR team, and it would be great to see some of it through. I expect to take up the fellowship at Queen Mary within the next half year or so. I’m very much looking forward to working with the C4DM folks again! Queen Mary&#…