A Sound Source Separation System for Percussion Instruments

Introduction:

We have developed a sound source separation system that extracts information such as the kind of instrument, onset time, and loudness of each note from an acoustic signal that consists of sounds of several kinds of percussion instruments. We proposed a new technique of sound source separation for percussion instruments. The technique, an improved version of template matching, enables the system to recognize each sound of percussion instruments even if the input is the mixture of several sounds or if the loudness varies. Experimental results show that this system can extract all notes from the sounds of some 8 beat drum patterns.

The system can detect the onset times of each drum sound in a popular-music drum performance, which is a polyphonic sound mixture of nine kinds of drum instruments: the bass drum (BD), snare drum (SD), low tom (LT), middle tom (MT), high tom (HT), hihat close (HC), hihat open (HO), ride cymbal (RI), and crash cymbal (CR). As far as we know, this system developed in 1992 and 1993 was the earliest work that can deal with such polyphonic drum-sound mixture, while the system cannot deal with a sound mixture of both drums and other musical instruments. The system is based on a template-matching method that is extended to identify a template pattern in a sound mixture by introducing a new distance measure. This measure can compute an appropriate distance between a pattern in the input power spectrum and the corresponding template pattern even though the input is a mixture of several template patterns. Each template pattern corresponds to a different drum sound: we prepared nine template patterns for the nine kinds of drum instruments.

Experimental Results:

The following figures are excerpts of our experimental results described in the reference [1] (published in a Japanese journal named ``The Transactions of the Institute of Electronics, Information and Communication Engineers D-II''). By clicking a figure, you can see the large (original) version.

Figure 1: (Experiment 1) An original score. The input audio signal generated by a MIDI sound module was a performance of this score represented by a standard MIDI file (SMF). The tempo was 120 M.M.


Figure 2: (Experiment 1) The correct answer (the ground truth). This corresponds to the above original score: this is yet another representation of the score.


Figure 3: (Experiment 1) The output of our system. Each row corresponds to a different drum instrument and triangle marks on each row shows its onset times detected by the system. The horizontal width of a triangle mark indicates the amplitude. A small vertical line segment in the row of the hihat open (HO) indicates the offset time, the time when it stops sounding.


Figure 4: (Experiment 2) An original score. The input audio signal generated by a MIDI sound module was a performance of this score represented by a standard MIDI file (SMF). The tempo was 120 M.M.


Figure 5: (Experiment 2) The correct answer (the ground truth). This corresponds to the above original score: this is yet another representation of the score.


Figure 6: (Experiment 2) The output of our system. Each row corresponds to a different drum instrument and triangle marks on each row shows its onset times detected by the system. The horizontal width of a triangle mark indicates the amplitude. A small vertical line segment in the row of the hihat open (HO) indicates the offset time, the time when it stops sounding.


Figure 7: A template pattern of a snare-drum (SD) sound. Each template pattern is represented as the power distribution that consists of frequency components of its sound. Those frequency components are extracted from the FFT-based power spectrum.


Figure 8: An example of the power distribution of the input sound mixture. Our method finds a template pattern in this distribution even if a pattern is overlapped with other patterns and the amplitude of a pattern is different from the original template pattern. This representation of the power distribution with the log-scale frequency axis in cents ([cent]) is used for detecting the bass drum (BD), snare drum (SD), low tom (LT), middle tom (MT), and high tom (HT). In our formulation, there are 100 cents to a tempered semitone and 1200 to an octave.


Figure 9: Another representation of the power distribution that is same with Figure 8. This representation of the power distribution with the linear frequency axis in hertz ([Hz]) is used for detecting the hihat close (HC), hihat open (HO), ride cymbal (RI), and crash cymbal (CR).

References:

1. Masataka Goto and Yoichi Muraoka: A Sound Source Separation System for Percussion Instruments, The Transactions of the Institute of Electronics, Information and Communication Engineers D-II, Vol.J77-D-II, No.5, pp.901-911, May 1994 (in Japanese).
2. Masataka Goto, Masahiro Tabuchi and Yoichi Muraoka: An Automatic Transcription System for Percussion Instruments, Proceedings of the 46th Annual Convention IPS Japan, 7Q-2, March 1993 (in Japanese).

Back to:

• Masataka Goto's Home Page