In this paper, we propose a method for supervised single-channel speech separation through sparse decomposition using periodic signal models. The proposed separation method employs sparse decomposition, which decomposes a signal into a set of periodic signals under a sparsity penalty. In order to achieve separation through sparse decomposition, the decomposed periodic signals have to be assigned to the corresponding sources. For the assignment of the periodic signal, we introduce clustering using a K-means algorithm to group the decomposed periodic signals into as many clusters as the number of speakers. After the clustering, each cluster is assigned to its corresponding speaker using preliminarily learnt codebooks. Through separation experiments, we compare our method with MaxVQ, which performs separation on the frequency spectrum domain. The experimental results in terms of signal-to-distortion ratio show that the proposed sparse decomposition method is comparable to the frequency domain approach and has less computational costs for assignment of speech components.

A fast and accurate method for Generalized Harmonic Analysis is proposed. The proposed method estimates the parameters of a sinusoid and subtracts it from a target signal one by one. The frequency of the sinusoid is estimated around a peak of Fourier spectrum using binary search. The binary search can control the trade-off between the frequency accuracy and the computation time. The amplitude and the phase are estimated to minimize the squared sum of the residue after extraction of estimated sinusoids from the target signal. Sinusoid parameters are recalculated to reduce errors introduced by the peak detection using windowed Discrete-Time Fourier Transform. Audio signals are analyzed by the proposed method, which confirms the accuracy compared to existing methods. The proposed algorithm has high degree of concurrency and is suitable to be implemented on Graphical Processing Unit (GPU). The computational throughput can be made higher than the input audio signal rate.

In this article, we considered the asymptotic expectations of the prediction error and the fitting error of a regression model, in which the component functions are chosen from a finite set of orthogonal functions. Under the least squares estimation, we showed that the asymptotic bias in estimating the prediction error based on the fitting error includes the true number of components, which is essentially unknown in practical applications. On the other hand, under a suitable shrinkage method, we showed that an asymptotically unbiased estimate of the prediction error is given by the fitting error plus a known term except the noise variance.

In the problem of determining the major frequency components of a signal disturbed by noise, a model selection criterion has been proposed. In this paper, the criterion has been extended to cover a penalized cost function that yields a componentwise shrinkage estimator, and it exhibited a consistent model selection when the proposed criterion was used. Then, a simple numerical simulation was conducted, and it was found that the proposed criterion with an empirically estimated componentwise shrinkage estimator outperforms the original criterion.