The DiTME Project
Interdisciplinary research in music technology
4.2 Onset detection system based on comb filters
The ODTW provides a remarkable improvement on detecting the slow onset of the tin whistle. Nevertheless, problems of strong amplitude and frequency modulations are still present in the ODTW system. However, these limitations are overcome by a technique for detecting note onsets using FIR comb filters which have different filter delays (Gainza et al. 2005). Figure 12, a block diagram illustrating the different components of the comb filters system is depicted.
The onset detector focuses on the harmonic characteristics of the signal, which are calculated relative to the energy of the frame. Both properties are combined by utilising FIR comb filters on a frame-by-frame basis. In order to generate an onset detection function the changes of the signal harmonicity are tracked. This produces peaks in the harmonicity changes that a new onset provides in the signal.
The method relates the harmonicity detection to the energy of the analysing frame, which is suitable for detecting slow onsets, and provides an accurate onset estimation time. The approach is robust for dealing with amplitude modulations: if the energy of the signal changes between successive frames (but not its harmonicity) the onset detection function remains stable. In addition, the method is robust to frequency modulations that gradually occur in the signal, since the signal harmonicity does not change considerably between frames.
Apart from amplitude modulations, frequency modulations can also arise in the signal, which consequently affect the onset detection accuracy. In, Figure 13, the onset detection function of a tin whistle signal playing E5 is depicted in the bottom plot. The middle and top plots depict the waveform and the spectrogram of the tin whistle signal respectively, where the amplitude and frequency modulations that arise in the signal can be seen. The E5 note depicted in, Figure 13, is played using a slide effect, which inflects the pitch to reach F5#, which means that a modulation between approximately 659 Hz to 740 Hz has occurred, Figure 13.
The onset detection function of, Figure 13, depicts very distinctive peaks at the position of the onsets. It can also be seen that the slide effect does not alter the accuracy of the detection. The onset detector has been evaluated by using two different databases, which comprise tin whistle tunes and other Irish traditional music instrument tunes respectively. The results show a clear improvement upon comparison with existing onset detection approaches.