Research On Resonance State Of Singing Voice Signal

Bel canto is a kind of graceful resonance singing style which is different from other vocal music genres. So many scholars have been focusing their researches on finding the difference of vocalization method and spectral energy distribution between bel canto and other vocal music genre. In this paper, the timbre research has been conducted based on the source-filter model. The features of time domain, frequency domain, and quefrency domain are extracted, and they are used in the classifications of bel canto singing voices and speaking voices, and amateur singers classification. The specific work includes:Six spectrum envelope extraction methods, the sinusoidal model, and the harmonic plus noise model are used to perform speech synthesis experiments. The cepstrum methods not only extract spectrum envelope that fits the spectrum well, but also have higher Mean Opinion Score (MOS) in the synthesized signals. The signals synthesized with harmonic plus noise model have the best timbre, and timbre of the signals synthesized with the sinusoidal model is better than that of the six spectrum envelope methods. Sounds synthesized with other methods also preserve most of their original timbre information. The spectrum envelope contains information of timbre and reflects the difference of resonance methods of different phonation ways.The three scalar characteristics of time domain:jitter, shimmer, and noise-to-harmonic ratio are used to measure the ability to the vocal cord of the singer. It is found that jitter gives better classification than that of shimmer and noise-to-harmonic ratio in the classification experiments of singing voice and speaking voice recorded by bel canto students. Bel canto students can better control the vocal cord vibration frequency while singing. One vector and four scalar characteristics of frequency domain:long-term average spectrum, singing power ratio, singer’s formant frequency, spectrum slope, and power ratio between low-high frequency components are also examed in the experiments. It is found that bel canto and male singers have larger singer power ratio. Also, bel canto songs have larger spectrum slope and low-high component power ratio. The feature of singer’s formant frequency performs badly in bel canto, pop and reading samples classification, whilst other frequency domain features perform not too bad. Four vector characteristics of quefrency domain:cepstrum coefficient, Mel Frequency Cepstrum Coefficient (MFCC), harmonic-suppressed cepstrum coefficient, and harmonic-suppressed MFCC are compared via classification experiments as well. It is found that vectors with larger dimension won’t definitely give better classification and the best classification dimension changes in different experiments.Among the three kinds of features, quefrency domain features perform best in the classification, especially the harmonic-suppressed cepstrum coefficient and MFCC. It is noticed that bel canto signals have larger energy in the vicinity of3kHz and smaller energy in the high frequency domain beyond5kHz in the long-term average spectrum. And this is the basis of the classification by frequency domain and quefrency domain features of different kinds of voice samples. The feature combination is also checked in the classification experiments. Though no combination brings any improvement compared to the best result of single feature classification, some combinations do give the better classification. Features combing is a good attempt to improve the classification result.