Abstract
A typical music band is composed of a vocal, an electric guitarist, an electric bassist, a drummer, and a keyboardist. The task of a keyboardist is to utilize the music instruments plugged-in in a keyboard appropriately. Nevertheless, keyboard sheets are hard to obtain. A keyboard beginner usually refers to guitar tabs to practice, thus the information of the instruments decision is lost. In this thesis, we have built a system of classification in an attempt to solve this problem. Each music instrument family data is composed of various pitches in 1 second. Also, duo-timbre and trio-timbre are mixed in order to generate mixtures and they serve as different labels. Their feature vectors are composed of a low-pass filtered power spectrogram, a high-pass filtered power spectrogram, a chromagram, and the time domain waveform. Several machine learning methods have been applied respectively, yet not all of the methods perform well. The k-nearest neighbors method has the most accurate result in both validation step (71.1%) and testing step (65.2%). We also have carried out a hearing test in order to understand whether the ability of classification for humans can compete with computers. As a result, humans’ accuracy is lower than computers’ in average.