The variational autoencoder (VAE) is a powerful generative model of speech signal. In the problem of speech enhancement, the VAE is used to model the clean speech whereas the non-negative matrix factorization is used to model the noise. The Wiener filter can be applied to extract the clean speech from the mixture signal. In this paper, we extend this approach by proposing a more sophisticated VAE framework, so-called disentangled recurrent variational autoencoder (RVAE), to model the clean speech. It takes the strategy of pre-training plus fine-tuning. Firstly, the clean speech and noise are pre-trained using two dependent RVAEs. Secondly, by transferring the knowledge learned in these two RVAEs, the third RVAE is used to train the mixture signal, resulting in disentanglement of the latent variables of the clean speech and noise. Experimental results demonstrate that the proposed approach outperforms other VAE-based approaches in terms of the signal-to-distortion ratio.

This article proposes a hybrid reverberation model by integrating two conventional models, namely, the multichannel linear prediction (MCLP) model and the spatial coherence model. The late reverberation is divided into two components. One component is modeled using an MCLP model, and the other is modeled using the spatial coherence model. In contrast with the conventional models, the proposed hybrid model increases modeling capacity, especially in the case of long reverberation time. In order to optimally estimate model parameters, joint speech dereverberation and separation is taken into account. The hybrid reverberation model is then used in conjunction with the multichannel nonnegative matrix factorization (MNMF). The method called Hybrid-FastMNMF is proposed by treating the reverberation component modeled by the spatial coherence model as a noise source and estimating its parameters similarly to speech sources. Furthermore, prior knowledge of the spatial coherence matrix is employed to whiten the observations, resulting in another method called Hybrid-FastMNMF-W. Experimental findings demonstrate the proposed methods' superior performance in terms of joint speech dereverberation and separation, and they further justify the efficiency of the proposed hybrid reverberation model.

In the problem of target tracking, different types of biases can enter into the measurement collected by sensors due to various reasons. In order to accurately track the target, it is essential to estimate and correct the measurement bias. Considering practical backgrounds, the bias is assumed to be locally stationary Gaussian distributed and an iterative estimation algorithm is proposed. Firstly, a mechanism is established to detect whether the bias switches between different Gaussian distributions. Secondly, the expectation maximization algorithm with the assistance of extended Kalman filtering and smoothing is proposed to iteratively estimate the bias and target state in an offline manner. Simulations show the proposed algorithm can suppress the impact of the measurement bias on target tracking.

In target tracking, the measurements collected by sensors can be biased in some real scenarios, e.g., due to systematic error. To accurately estimate the target trajectory, it is essential that the measurement bias be identified in the first place. We investigate the iterative bias estimation process based on the expectation-maximization (EM) algorithm, for cases where sufficiently large numbers of measurements are at hand. With the assistance of extended Kalman filtering and smoothing, we derive two EM estimation processes to estimate the measurement bias which is formulated as a random variable in one state-space model and a constant value in another. More importantly, we theoretically derive the global convergence result of the EM-based measurement bias estimation and reveal the link between the two proposed EM estimation processes in the respective state-space models. It is found that the bias estimate in the second state-space model is more accurate and of less complexity. Furthermore, the EM-based iterative estimation converges faster in the second state-space model than in the first one. As a byproduct, the target trajectory can be simultaneously estimated with the measurement bias, after processing a batch of measurements. These results are confirmed by our simulations.

We investigate the problem of target tracking using a wireless sensor network with asynchronous sensors. To study the impact of sensor clock imperfection on target tracking in practical situations, we build a testbed and collect data from an outdoor experiment. After analyzing the collected data, we find that the TDOA (time‐difference‐of‐arrival) and FDOA (frequency‐difference‐of‐arrival) measurements have notable bias, which is caused by asynchronous sensors or more precisely by the sensor clock drift. Based on the model of clock drift, the measurement bias and the target position are integrated into a state‐space model. Both can be estimated in the framework of the extend Kalman filter. In some circumstance, the target trajectory is tracked successfully.

The problem of source enumeration in array processing is investigated. In an information theoretic criterion framework, we use in addition to the probability density function of observations, the probability density function of the sample eigenvalues obtained from the sample covariance matrix of the observations. Although the latter adds information to the criterion it is widely ignored by most traditional approaches. Simulations show that the significant performance gain offered by the proposed criterion in terms of correctly detecting the number of sources in some difficult situations, such as small sample sizes, low signal-to-noise power ratio, close spacing and high correlation between sources.

We consider the problem of source number estimation in array processing when impulsive noise is present. To combat impulsive noise more effectively, two robust estimators with high breakdown points, i.e., the minimum covariance determinant (MCD) estimator and the MM-estimator are applied in combination with the bootstrap. The MCD estimator is applied to discard the outliers in observations caused by impulsive noise, while the MM-estimator is suggested to estimate robustly the covariance matrix of observations. Simulations show the significant performance gain offered by the two proposed methods in terms of correctly estimating the number of sources at low SNR in the presence of impulsive noise.

In this paper bootstrap resampling techniques are applied to assess speech quality and thereby evaluate performance of distinct speech enhancement algorithms, under the assumption that the speech segments can be approximated by an autoregressive model. A bootstrap-based multiple hypotheses testing procedure is constructed to test a distance measure based on linear predictive coding, which is the log-likelihood ratio distance. It is shown that the multiple hypotheses test results correlate well with conventional numerical distance measures, which suggests the applicability of the proposed procedure in assessment of speech quality as well as speech enhancement algorithms.