(A) study on acoustic sound separation, timbre recognition and unwanted audio cancellation for multimedia acoustic systems

Abstract

In this paper we show signal processing techniques for multimedia acoustic system. First, we show the acoustic source separation technique using multiple microphone based on the sinusoid modeling. Until now there are many other approaches for separating acoustic sources, such as acoustic beamforming, auditory scene analysis and blind deconvolution method. In this thesis the sound recorded from multiple microphones is decomposed into several sinusoids using proposed sequential sinusoid estimation method. We estimate many sinusoids one by one, and each sinusoid parameters are estimated using the steepest descent estimation methods. The accuracy of estimation is better than previous sinusoid decomposition method. After decomposing sinusoids we classify them into dedicated groups which belongs to specific sound sources. The proposed separation technique is more robust than the wideband beamforming method, and show better separation quality than blind deconvolution method. We can use the proposed technique even in reverberant acoustic channel environment and even under-determined case (the number of sound source is bigger than the number of microphones). The limitation of the proposed method is the residual which is not modeled with sinusoidal modeling. Second, we show an instrument timbre recognition method based on the hidden Markov model (HMM). Sinusoid model is also used in this technique. Spectral envelop is the key information of music instrument timbre. We decompose recorded instrument sound into sinusoidal components (harmonics) and noise component, and estimate the amplitudes of harmonics component. We express spectral envelop effectively using amplitudes of estimated sinusoids. Three kinds of features are used to apply recognition procedure, and continuous single Gaussian output HMM is used. To evaluate the performance of recognition method, real instrumental sound recorded from MUMS (MacGill University Master Samples) is used to recognize the timber of in...