Saman Mousazadeh

Voice Activity Detection in Presence of Transient Noise Using Spectral Clustering

Voice activity detection has attracted significant research efforts in the last two decades. Despite much progress in designing voice activity detectors, voice activity detection (VAD) in presence of transient noise is a challenging problem. In this paper, we develop a novel VAD algorithm based on spectral clustering methods. We propose a VAD technique which is a supervised learning algorithm. This algorithm divides the input signal into two separate clusters (i.e., speech presence and speech absence frames). We use labeled data in order to adjust the parameters of the kernel used in spectral clustering methods for computing the similarity matrix. The parameters obtained in the training stage together with the eigenvectors of the normalized Laplacian of the similarity matrix and Gaussian mixture model (GMM) are utilized to compute the likelihood ratio needed for voice activity detection. Simulation results demonstrate the advantage of the proposed method compared to conventional statistical model-based VAD algorithms in presence of transient noise.

AR-GARCH in Presence of Noise: Parameter Estimation and Its Application to Voice Activity Detection

This paper presents a new method for voice activity detection (VAD) based on the autoregressive-generalized autoregressive conditional heteroscedasticity (AR-GARCH) model. The speech signal is modeled as an AR-GARCH process in the time domain, and the likelihood ratio is computed and compared to a threshold. The time-varying variance of the speech signal needed for computing the likelihood function under speech presence hypothesis, is estimated using the AR-GARCH model. The model parameters are estimated using a novel technique based on the recursive maximum likelihood (RML) estimation. The variance of the additive noise, a critical issue in designing a VAD, is estimated using the improved minima controlled recursive averaging (IMCRA) method, which is properly modified to be applicable to noise variance estimation in the time domain. The performances of the VAD and the parameter estimation method are examined under several conditions. Experimental results indicate the robustness of the AR-GARCH based VAD both to noise variations and low signal-to-noise ratio (SNR) conditions.

Embedding and function extension on directed graph

In this paper, we propose a novel technique for finding the graph embedding and function extension for directed graphs. We assume that the data points are sampled from a manifold and the similarity between the points is given by an asymmetric kernel. We provide a graph embedding algorithm which is motivated by Laplacian type operator on manifold. We also introduce a Nystrom type eigenfunctions extension which is used both for extending the embedding to new data points and to extend an empirical function on new data set. For extending the eigenfunctions to new points, we assume that only the distances of the new points from the labelled data are given. Simulation results demonstrate the performance of the proposed method in recovering the geometry of data and extending a function on new data points. HighlightsA novel technique for finding the graph embedding and function extension for directed graphs is proposed.The data points are assumed to be sampled from a manifold and the similarity between the points is given by an asymmetric kernel.A graph embedding algorithm motivated by Laplacian type operator on manifold is proposed.A Nystrom type eigenfunctions extension for extending the embedding an empirical function on new data set is proposed.In extension phase, we assume that only the distances of the new points to the labelled data are given.

Simultaneous parameter estimation and state smoothing of complex GARCH process in the presence of additive noise

ARCH and GARCH models have been used recently in model-based signal processing applications, such as speech and sonar signal processing. In these applications, additive noise is often inevitable. Conventional methods for parameter estimation of ARCH and GARCH processes assume that the data are clean. The parameter estimation performance degrades greatly when the measurements are noisy. In this paper, we propose a new method for parameter estimation and state smoothing of complex GARCH processes in the presence of additive noise. Simulation results show the advantage of the proposed method in noisy environments.

Anomaly detection in sonar images based on wavelet domain noncausal AR-ARCH random field modeling

In this paper we introduce a novel anomaly detection method in sonar images based on noncausal autoregressive-autoregressive conditional heteroscedasticity (AR-ARCH) model. The background of the sonar image in the wavelet domain is modeled by a noncausal AR-ARCH model. Matched subspace detector (MFD) is used for detecting the anomaly in the image. The proposed method is computationally efficient and is robust to the orientation variation of the image, compared to competing method.

Two dimensional noncausal AR-ARCH model: Stationary conditions, parameter estimation and its application to anomaly detection

Image anomaly detection is the process of extracting a small number of clustered pixels which are different from the background. The type of image, its characteristics and the type of anomalies depend on the application at hand. In this paper, we introduce a new statistical model called noncausal autoregressive-autoregressive conditional heteroscedasticity (AR-ARCH) model for background in sonar images. Based on this background model, we propose a novel anomaly detection technique in sonar images. This new statistical model (i.e. noncausal ARCH) is an extension of the conventional ARCH model. We provide sufficient stationarity conditions and develop a computationally efficient method for estimating the model parameters which reduces to solving two sets of linear equations. We show that this estimator is asymptotically consistent. Using matched subspace detector (MSD) along with noncausal AR-ARCH modeling of the background in the wavelet domain, we propose an anomaly detection algorithm for sonar images, which is computationally efficient and less dependent on the image orientation. Simulation results demonstrate the performance of the proposed parameter estimation and the anomaly detection algorithm.

Multisensory speech enhancement in noisy environments using bone-conducted and air-conducted microphones

In this paper, we propose a speech enhancement algorithm for estimating the clean speech using samples of air-conducted and bone-conducted speech signals. We introduce a model in a supervised learning framework by approximating a mapping from concatenation of noisy air-conducted and bone-conducted speech to clean speech in the short time Fourier transform domain. Two function extension schemes are utilized: geometric harmonics and Laplacian pyramid. Performances obtained from the two schemes are evaluated and compared in terms of spectrograms and log spectral distance measures.

Voice activity detection in presence of transient noise using spectral clustering and diffusion kernels

In this paper, we introduce a voice activity detection (VAD) algorithm based on spectral clustering and diffusion kernels. The proposed algorithm is a supervised learning algorithm comprising of learning and testing stages: A sample cloud is produced for every signal frame by utilizing a moving window. Mel-frequency cepstrum coefficients (MFCCs) are then calculated for every sample in the cloud in order to produce an MFCC matrix and subsequently a covariance matrix for every frame. Utilizing the covariance matrix, we calculate a similarity matrix using spectral clustering and diffusion kernels methods. Using the similarity matrix, we cluster the data and transform it to a new space where each point is labeled as speech or nonspeech. We then use a Gaussian Mixture Model (GMM) in order to build a statistical model for labeling data as speech or nonspeech. Simulation results demonstrate its advantages compared to a recent VAD algorithm.

Out-of-sample extension of band-limited functions on homogeneous manifolds using diffusion maps

In this paper, we address the problem of function extension when the available data lies on a homogeneous manifold (i.e. the domain of the function is a homogeneous manifold embedded in the Euclidean space) and the function is band-limited. We solve this problem in the general case in which the manifold is unknown. We assume that we have sufficient labeled data to reconstruct the function from labeled data. We also assume that we have enough data (at least exponential in the intrinsic dimension of the manifold) to approximate the Laplace-Beltrami operator on the manifold. The proposed method has a closed form solution and consists of matrix multiplication and inversion. As the size of data approaches infinity, the proposed method converges to the optimal solution as long as the function values are known on an appropriate sampling set. Simulation results demonstrate the advantage of the proposed method over commonly used function extension methods. HighlightsWe introduced a new statistical new supervised learning method.We assumed that data lies on a homogenous manifold.We used diffusion maps as a manifold learning algorithm.We used this method to estimate the parameters of ARMA process and the controlling parameters of acoustic impulse response.Simulation results show the performance of the proposed method.

Voice activity detection in transient noise environment using Laplacian pyramid algorithm

Voice activity detection (VAD) has attracted significant research efforts in the last two decades. Despite much progress in designing voice activity detectors, voice activity detection in presence of transient noise and low SNR is a challenging problem. In this paper, we propose a new VAD algorithm based on supervised learning. Our method employs Laplacian pyramid algorithm as a tool for function extension. We estimate the likelihood ratio function of unlabeled data, by extending the likelihood ratios obtained from the labeled data. Simulation results demonstrate the advantages of the proposed method in transient noise environments over conventional statistical methods.