Timbre Recognition Of Western Instruments

Musical instrument recognition is one of the most important issues in music information retrieval field. The main objective of musical instrument recognition is to identify the family and name of musical instrument from the given sound, which can be achieved by the analysis of timbre. Although a large number of features have been used for classification in the existed algorithms, the essence of timbre was often ignored. Focusing on the information of timbre, this thesis studies the recognition of western instruments.The appropriate mathematical model is an important basis of timbre recognition. In this thesis, three types of musical signal models are summarized and the common timbre features are introduced. Based on the instrument sounding mechanism, the source-filter model is heavily reconsidered, where musical sound is modeled as a convolution between excitation and resonator. The resonator contributes greatly to the timbre and its information can be analyzed by cepstrum coefficients. Combines the instrument sounding mechanism with the human auditory perception process, Mel Frequency Cepstral Coefficient (MFCC) is an outstanding timbre feature in the quefrency domain.The synthesizing model is built up in this thesis, which can describe the musical sound with more details. As an improvement of source-filter model, the excitation part is modeled as sinusoids plus noise. Based on the synthesizing model, the vector feature of Nontonal Mel Frequency Cepstral Coefficient (NMFCC) is proposed as a new timbre aspect. This feature is derived from the nontonal spectrum and offers an exact description of resonator through MFCC. With the Empirical Mode Decomposition (EMD), the musical sound can be decomposed into components called Intrinsic Mode Function (IMF). Another proposed timbre feature in this thesis is Mel Frequency Cepstral Coefficient of Low Order Modes Summation (LOMS-MFCC), which highlights the formants in high frequency range.Through the massive classification experiments, the effectiveness of the new timbre features NMFCC and LOMS-MFCC is proved. Even compared with the outstanding feature MFCC, the proposed features also bring higher recognition rates. No matter which classifier is used, or what classification task is to be implemented, the NMFCC always shows prominent performance in the recognition of western instruments. Based on this excellent timbre feature, a better result is acquired when the proper time domain features are combined in instrument recognition system.