Abstract
This paper presents a new algorithm for automatic transcription of drums by the adaptive processing method. In transcribing performed drum music, a system must extract each beat and must recognize the drum's type at each beat.
In order to extract a drum's beating time based on time-domain, we employed recursive estimation technique which should make transient change extraction more accurately and eliminate the necessity of frequency analysis. The approach presented here utilized the normalized LSL algorithm with the likelihood variable estimation, including the threshold processing and local method. This recursive algorithm updates its representation at every music signal sample using exponentially weighted past data. Thus it is possible to track the beating changes in the drum music signal. From the extracted beating times, we modeled the drum music nonstationary signal by an AR model with the time-varying coefficients and estimated the time-varying coefficients using the RLS adaptive algorithm to obtain a time-frequency spectrum of the drum signal. With the obtained time-frequency spectrum, the drum's type recognition can be carried out by the pattern matching method. Finally, we combined the beating time extraction and the drum's type recognition to form an automatic transcribing system of the drum music, and presented experiment results of transcription to illustrate a validity of the proposed system for the drum machine data and the real drum music data.
