Most of my current research is in the area of Independent Component Analysis (ICA) and related techniques, applied to audio and music signals. Specific research areas include: Analysis and transcription of musical audio signals using ICA and related techniques; Object-based coding of audio; Rhythm tracking and automatic accompaniment; Links between ICA and efficient information transmission; and Non-negative ICA. This work takes place in the Centre for Digital Music, within the Digital Signal Processing & Multimedia Group.

Blind Source Separation and Independent Component Analysis

My research is in the area of signal processing known as blind source separation (BSS) using independent component analysis (ICA). Sometimes known as the "cocktail party problem", the task is to recover original source signals from observations of mixtures. This problem is found many real-world signals, and these methods are finding increasing application in the analysis of e.g. EEG signals, audio, satellite images, and text documents. In the Centre for Digital Music I am investigating theory and algorithms for BSS and ICA, and applying these to the analysis of musical audio signals. This has proved to be a very exciting and fruitful area: its importance is indicated for example by the widespread interest in our EPSRC-funded ICA Research Network (initially with 44 researchers in 26 UK institutions), and success of the recent international ICA conferences.

I have developed a Non-negative ICA approach to separation of positive sources (Plumbley, 2001): many real world tasks (including music spectra) share these non-negative features. I have proved sufficient conditions for algorithms to find this non-negative ICA solution (Plumbley, 2002), and developed and analyzed novel algorithms for non-negative ICA with Prof. Erkki Oja (Plumbley, 2003; Plumbley & Oja, 2004). This work led me to investigate the geometrical structure of the algorithm search space (constrained to orthonormal matrices) in terms of a Lie group structure. Using this geometric view gives a deep insight into methods for non-negative ICA and “standard” ICA methods, allowing faster searching (Plumbley 2005 [Neurocomputing]).

Automatic Music Transcription and Sparse Coding of Music

I have investigated the application of ICA, and the related method of sparse coding, to the challenging task of automatic music transcription, the extraction of notes played by a musical instrument from a musical audio recording. Sparse coding, where it is assumed that few sources are active at any time, is good model for many real-world problems. In music, for example, few notes are assumed to be on at any one time. On my EPSRC grant “Automatic Polyphonic Music Transcription Using Multiple Cause Models and Independent Component Analysis” we developed a novel non-negative sparse coding method to perform this transcription, in an unsupervised manner, without prior information about notes spectra (Abdallah & Plumbley, 2005). One of the Assessors of the Final Report of this grant considered the demonstrated results to be “nothing short of amazing” and that “it is a major step forward for this technology”. The overall assessment of this project by ESPRC was “Outstanding”.

Beat Tracking and Dynamics of Musical Audio

I am currently investigating methods to analyze the temporal structure of music, such as onset detection and beat tracking, with one of my PhD students, Matthew Davies (Davies & Plumbley 2004, 2005), and we will undertake a fundamental investigation of the temporal structure of musical signals on the recently awarded “Information Dynamics of Music” EPSRC project. Reviewers commented that “this kind of research is badly needed” and is of “fundamental importance”.

Object-Based Coding of Musical Audio

I am building on the music analysis and transcription work with Emmanuel Vincent on the EPSRC grant on Object-Based Coding of Musical Audio, following on from work of my PhD student Paul Brossier. Here we aim to use the analysis, transcription and modelling of musical audio to produce compact and scalable “object-based” representations of music which could be used to transmit music at very low bit rates, much lower than the ubiquitous “MP3” format, for example. As recognized by one of the grant reviewers, this project “tackles crucial and as yet unresolved issues” in audio coding.

Recent Research Funding


  • Sparse Representations for Signal Processing and Coding (EPSRC Grant EP/D000246/1, £282,902, 2005-2008), with Mike Davies.
  • Techniques and Algorithms for Understanding the Information Dynamics of Music (EPSRC Grant GR/S82213/01, £184,965 to Queen Mary, 2005-2007), with Mark Sandler at Queen Mary, Geraint Wiggins and Michael Casey at Goldsmiths College, London.
  • ICA Research Network, [Blind Source Separation and Independent Component Analysis Network] (EPSRC Grant EP/C005554/1, £62,058, 2005-2007), with Mike Davies at Queen Mary, plus 44 other researchers at 26 other UK institutions.
  • Object-based Coding of Musical Audio (EPSRC Grant GR/S75802/01, £202,353 , 2004-2007), with Mark Sandler and Mike Davies.
  • Advanced Subband Systems for Audio Source Separation (EPSRC Grant GR/S85900/01, £176,431, 2004-2007), with Mike Davies and Mark Sandler.
  • SIMAC: Semantic Interaction with Music Audio Contents (EU Grant FP6-IST-507142, €2,982,921 total, £219,000 to Queen Mary, 2004-2006), with Mark Sandler at Queen Mary. Partners: UPF (Spain), ÖFAI (Austria), Matrix Data (UK), Philips Research (Netherlands).
  • Digital Music Research Network []. (EPSRC Grant GR/R64810/01, "Music Processing Network" £61,794, 2003-2005), with Mike Davies and Mark Sandler at Queen Mary, plus researchers at six other UK institutions.