自動採譜のための楽器音同定及び音高推定手法

抄録

In this paper, a musical instrument and pitch estimation method of polyphonic music signals for automatic transcription is proposed. The method consists of three main stages. First, spectrum of polyphonic musical signal is calculated to extract audio features using short-time FFT (Fast Fourier Transform). Next, the polyphonic music signal is decomposed to monophonic music signals based on HC (Harmonic Clustering) in log scale frequency domain. Finally, musical instrument of the signals is identified with tone database using GMM (Gaussian Mixture Model) or SVM (Support Vector Machine) applied to MFCC (Mel-Frequency Cepstrum Coefficient). Pitch is sequentially estimated using the extracted power spectrum peaks information with pitch database. Simulation results are given to demonstrate the effectiveness of tone and pitch estimation of polyphonic music signal with various combinations of musical instruments using RWC (Real World Computing) music database system.