Timbre Feature Analysis Of Musical Instruments

The research of classifying musical instruments is the foundation of exploring the nature of musical instruments. Due to the difference of musical instruments sounding mechanisms, the uncertainty of measuring timbre, and the limitation of human cognizing the process of auditory perception, the research of timbre is still in the bottleneck stage. The features extracted respectively from the time domain, frequency domain, and quefrency domain are studied in this paper.In the time domain, the envelope of single note is studied and the root-mean-square(RMS) envelope of a single note is cut into five segments. The ratio between the length of each segment to the length of the note ratio is calculated as a feature in the time domain, which is called as envelope segment length ratio. Simulation results show that this ratio is good at classifying musical instruments within one kind of instrument family.in frequency domain, three kinds of spectra are obtained from short time Fourier transform (STFT), constant Q transform (CQT) and modified constant Q transform (MCQT). respectively. The means and standard deviations of spectral width, spectral asymmetry, spectral kurtosis, spectral centroid, spectral roll-off and spectral flux are extracted from these three kinds of spectra. Simulation results show that the feature combination of CQT spectra is over all the best at classifying notes, and the feature combination of MCQT spectra has the best performance of distinguishing notes between different instruments in one kind of family.In quefrency domain, three kinds of cepstra are obtained from STFT, CQT and MCQT, respectively. Combined with hearing perception model, STFT based MFCC and CQT based MFCC is obtained. Compared the different cepstra, the effective parts for musical instrument classification are different. The experiments are done from two aspects including feature importance of every dimension and the similarity between different notes which is calculated from some dimension of feature in quefrency domain. By experiments, the proper dimensions in classification are the first25dimensions in STFT based cepstrum, the first7dimensions in CQT based cepstrum, the first12dimensions in MFCC and the first12dimensions in CQT based MFCC. The capacity difference of distinguish timbre between the high dimensional part and the low dimensional part is not obvious.Three kinds of spectra and five kinds of features in quefrency domain are used in the classification. The misclassification probability of STFT based MFCC is the smallest. The effect of classification is improved greatly by combining envelope segment length ratio with other feature. Classification will be more effective when envelope segment length ratios is combined with MCQT logrithmic spectrum or MCQT based cepstrum.