AdaMast: A Drum Sound Recognizer
based on Adaptation and Matching of Spectrogram Templates
We propose a template-matching-based system,
called AdaMast, that detects onset times of the bass drum,
snare drum, and hi-hat cymbals in polyphonic audio signals
of popular songs. AdaMast uses the power spectrograms
of the drum sounds as templates. However, there
are two main problems in transcribing drum sounds in the
presence of other sounds. The first problem is that actual
drum-sound spectrograms cannot be prepared as templates
beforehand for each song. The second problem is
that power spectrograms of sound mixtures including the
drum sound are greatly different from the template (pure
drum-sound spectrogram). To solve the first problem, a
template-adaptation algorithm is built into AdaMast. To
solve the second problem, a distance measure used in the
template matching is designed to be robust to the spectral
overlapping of other sounds. The test results in Audio
Drum Detection Contest were 72.8%, 70.2%, and 57.4%
in transcribing the bass drums, snare drums, and hi-hat
cymbals, respectively, and AdaMast won the contest.
Kazuyoshi Yoshii, Masataka Goto, Hiroshi G. Okuno:
"Drum Sound Recognition for Polyphonic Audio Signals by Adaptation
and Matching of Spectrogram Templates with Harmonic Structure Suppression,"
IEEE Transactions on Audio, Speech and Language Processing,
Vol.15, No.1, Jan. 2007.
This research was achieved by using
RWC Music Database.
We thank everyone who has made this database.
Author : Kazuyoshi Yoshii (AIST)
mail to k.yoshii(at)aist.go.jp