About Level3
Search archives
- June 2007
- August 2006
- May 2005
- June 2004
- November 2003
DIT Home

Read postings about this article   |   Post a comment about this article  |  print this article [pdf]

The DiTME Project

Interdisciplinary research in music technology


Author - Eugene Coyle, Dan Barry, Mikel Gainza, David Dorran, Charlie Pritchard, John Feeley and Derry Fitzgerald

[<<previous][ next>>]

Show/ hide article menu (click icons opposite)

3.3   Single channel source separation

The task of single channel source separation is significantly more difficult to achieve, nevertheless the DiTME team has given some consideration to the problem. In Barry et al. 2005 a method for detecting and extracting drums and other percussive signals from single channel music mixtures presented. The technique involves taking the first order log derivative of a short time Fourier transform. Following this, the number of positive tending bins are accumulated to form a percussive feature vector. The spectrogram is then modulated by this feature vector before resynthesis. Upon resynthesis only the percussive elements of the signal remain, Figure 10.

4.  Music transcription within Irish traditional music

Irish traditional music has passed from generation to generation largely by oral transmission: hence the lack of transcription of this valuable cultural heritage. In researching for his Ph.D. as a member of the DiTME team, Mikel Gainza made a number of significant contributions in digital signal processing techniques to provide an understanding of the nature of audio signals in traditional music performance. Traditional music is more monophonic in nature than classical or other forms of music. It may be played as a solo performance permitting the musician to express individual nuance in style and ornamentation, or in unison with other instruments. However, simplistic harmonic accompaniment has also been incorporated in recent years. In his Ph.D. thesis ‘Music Transcription within Irish Traditional Music’, Gainza has identified important features of recorded notes, in particular note onset detection characteristics associated with different traditional instrument types. The ‘slow’ onset characteristic of the tin whistle has been carefully analysed. Ornamentation and transcription in traditional music also features in Gainza’s research. In endeavouring to develop a robust automatic music transcription system, note feature characteristics must be understood. The ability to accurately detect note onset is particularly important as it provides an accurate means of recognising note commencement or event variation.

A review of existing onset detection methods in Gainza’s Ph.D. (2006) concludes that the main problems encountered by existing approaches are related to frequency and amplitude modulations, in fast passages such as legato, in the detection of slow onsets, and in detecting ornamentation events. A review of existing pitch detection methods was also undertaken in this thesis, which highlights that a system that detects the different types of ornamentation within Irish traditional music has not yet been implemented. In addition, the review shows that periodicity based methods are less accurate in application to polyphonic signals.

In order to overcome the problems identified in the literature review, different applications for onset, pitch and ornamentation detection are presented in Gainza’s research. These are summarised in sections 4.1 to 4.4.

4.1 Onset detection system applied to the tin whistle

 First an onset detection method which focuses on the characteristics of the tin whistle within Irish traditional music was developed. This is known as the Onset Detection System applied to the Tin Whistle (ODTW). (See Gainza et al. 2004a.) The different blocks of the proposed onset detector are depicted in, Figure 11.

A time-frequency analysis is first required, which splits the signal into different frequency bands. The energy envelope is calculated and smoothed for every band. Peaks greater than a band dependent threshold in the first derivative function of the smoothed energy envelope will be considered as onset candidates. Finally, all band peaks are combined to obtain the correct onset times.

The onset detection system utilises knowledge of the notes and modes that the tin whistle is more likely to produce, and the expected blowing pressure that a tin whistle produces per note. Problems arising in respect of legato playing in onset detection are catered for by utilising a multi-band decomposition, where one band is utilised per note. In an effort to reduce the effect of amplitude modulations, different novel thresholding methods have been implemented.

By using these methods in conjunction with an optimisation of other system parameters, the onset detection system deals with moderate signal amplitude modulations. A comparison was made of the ODTW against existing onset detection methods, configured with their respective best performing parameters: the ODTW has provided the best results.


[<<previous][ next>>