Timbre identification of instrumental music via energy distribution modeling

The traditional evaluation of instrumental music is generally based on experts. However, the expert-based evaluation strategy is seriously affected by a number of factors such as human's subjectivity and then decreases the evaluation reliability. This paper aims at automatically identifying the timbre of saxophone music, and proposes a new method based on the energy distribution in frequency domain of music signals. First, we transform music signals into frequency domain using short-time Fourier transformation (STFT). Then, we compute the spectral envelope, which may describe the rule of frequency attitude, based on linear predictive coding (LPC). At last, we find that the energy distribution can be approximated by an exponential function, and present a ONE-dimensional feature named average linearity value (ALV). The ALV feature measures the extent of closeness between the energy distribution and exponential functions and is used to distinguish high-level timbres from low-level timbres in our work. The experiment is conducted on 9 groups of data, and the experimental results demonstrate the effectiveness of this method.