Elizabeth Hoppe

Non-linear, adaptive array processing for acoustic interference suppression

A method is introduced where blind source separation of acoustical sources is combined with spatial processing to remove non-Gaussian, broadband interferers from space-time displays such as bearing track recorder displays. This differs from most standard techniques such as generalized sidelobe cancellers in that the separation of signals is not done spatially. The algorithm performance is compared to adaptive beamforming techniques such as minimum variance distortionless response beamforming. Simulations and experiments using two acoustic sources were used to verify the performance of the algorithm. Simulations were also used to determine the effectiveness of the algorithm under various signal to interference, signal to noise, and array geometry conditions. A voice activity detection algorithm was used to benchmark the performance of the source isolation.

Source localization performance of a multi-array network under sensor orientation and position uncertainty

There is increased interest in networking arrays of sensors for distributed source localization. An analysis of the generalized model and estimation performance for multiple sources being observed by a field of networked arrays has recently been studied in via the Cramer-Rao lower bound (CRLB). In previous work concerning the CRB and multiple sources, the models were formulated for observations with a distributed network of sensor arrays. However, previous work assumed completely known sensor orientation and position. In real world systems such as distributed sonar systems, this is rarely the case. In this paper, sensor orientation and position errors are incorporated into the signal model. Simulation examples are given that show that a network of a high number of low-complexity arrays outperforms a network of a low number of high resolution arrays when considering subarray position and orientation errors. This occurs even though the high resolution array network has four times as many sensing elements.

Principal component analysis for emergent acoustic signal detection with supporting simulation results

A method is introduced that uses principal component analysis (PCA) to detect emergent acoustic signals. Emergent signal detection is frequently used in radar applications to detect signals of interest in background clutter and in cognitive radio to detect the primary user in a frequency band. The method presented differs from other standard techniques in that the detection of the signal of interest is accomplished by detecting a change in the covariance between two channels of data instead of detecting the change in statistics of a single channel of data. For this paper, PCA is able to detect emergent acoustic signals by detecting when there is a change in the eigenvalue subspace of the covariance matrix caused by the addition of the signal of interest. The algorithms performance is compared to an energy detector and the Neyman-Pearson theorem. Acoustic simulations were used to verify the performance of the algorithm. Simulations were also used to examine the effectiveness of the algorithm under various signal-to-interferer and signal-to-noise ratios, and using various test signals.

A dual space approach to beam tracing for optimal microphone placement.

In many teleconferencing or video‐conferencing applications, a high degree of speech intelligibility is difficult to achieve due to sub‐optimal microphone placement in the room. In this work, beam tracing in the dual space is used to acoustically model an enclosed space with the goal of discovering optimal positions for microphone placement. In the majority of existing work, beam tracing is used to model empty rooms with simple layouts. This work allows for the addition of freestanding items, such as furniture or support pillars, and more complex room layouts. In addition, currently existing algorithms use beam tracing to predict and simulate the response of a source in a room at a specific location. In this work, both signals of interest as well as interferers will be included in the model. Instead of predicting the response at specific (x,y) locations, this work seeks to guide the placement of microphones for maximum speech intelligibility. The dual space representations for the reflected beams are gene...

Compressive sensing applied to audio signals using higher‐order projections and the spatial domain.

Compressive sensing (CS) is a new approach to data acquisition that is receiving much attention in the current literature. The main goal in CS is to accurately reconstruct a signal from a small number of randomly projected basis functions. The CS reconstruction accuracy hinges on the assumption that the signal can be projected into a sparse domain. A majority of the CS research to date has been focused on the application of CS to image processing. There has, however, been very limited research on the application of CS techniques to audio signals focused in two main areas: the use of the frequency and wavelet domains as the sparse signal domain, and the use of the spatial domain to perform compressive beamforming. In this work, two new approaches are examined. The first is the use of the spatial domain to reconstruct signals instead of simply identifying their direction of arrival. The second is the use of higher‐order projection techniques (such as principal component analysis) as a sparse domain. This work will apply these techniques to speech signals to examine the ability of CS to reconstruct wide‐band audio signals. In addition, the effect of additive noise on the reconstruction accuracy will be examined.

Distributed underwater source location estimation using a Multiarray Network

Even though there are four times as many total sensors, the two 32-channel array network had larger localization errors when compared to an eight two array network when variance was introduced into the position and orientation of the arrays. It was shown that the performance of either network was dependent on these variances, but the two 32-channel array network was much more sensitive to variance in position and orientation of its arrays for all source locations.

Non-linear, adaptive array processing for underwater source localization and sonar interference suppression

A method is introduced where blind source separation (BSS) of acoustical sources is combined with spatial processing to remove non-Gaussian, broadband interferers from space-time displays such as bearing track recorder (BTR) displays. This differs from standard techniques such as placing nulls in the direction of interferers. The algorithm performance is compared to adaptive beamforming techniques, such as minimum variance distortionless response beamforming (MVDR). Simulations and experiments using two acoustic sources were run to verify algorithm performance.

Distributed detection and estimation based on the turbo approach.

We develop a centralized information fusion architecture from basic principles of information theory and Bayesian statistics. It is well known that any clustering, quantizing, or thresholding of data causes loss of information unless a sufficient statistic is computed in the processing. For the case of wideband active ranging systems, the coherent output of an optimum beamformer and a matched filter is a sufficient statistic that can be transmitted to the fusion center. For unknown target velocity, range, and bearing, the wideband space‐time matched filter output can be interpreted as a multidimensional wavelet transform or a delay‐scale‐bearing map. In this paper, an iterative, Bayesian, joint estimation‐detection approach is used for computation of sufficient statistics and multisensor information fusion. An approach borrowed from sequential Bayesian processing is used to compute prior densities for joint Bayesian estimation‐detection. In this approach a posteriori densities become priors after a coordinate transformation that transforms the outputs of each sensor to a common reference frame for all sensors. In this paper, receiver operating characteristics and Cramer–Rao lower bounds are given for several undersea signal processing cases of interest.

Compact microphone array processing for in‐vehicle voice detection applications.

The main goal of voice activity detection algorithms is to determine the presence of human voice signals in a given environment. Voice activity detection is very challenging in vehicle interiors. The main challenge in detecting the presence of voice signals in vehicles is the presence of a large number of interferers and a high background noise level. Further, many types of interferers such as tire or engine noise have signals that are highly nonstationary. In this work, compact microphone arrays mounted in various locations within a vehicle are used to extract signals from locations of interest. Experimental comparisons of the performance of several voice activity detection algorithms are made for various array configurations (including single microphone) and source extraction processing algorithms. Processing algorithms considered include blind source separation algorithms such as fastICA, transfer function based inversion methods, and both fixed and adaptive beamforming techniques. The performance of c...

The role of reproducing densities in a joint likelihood approach to multi‐sensor detection and estimation

In this work, a Bayesian, joint estimation‐detection approach is used for computation of sufficient statistics and development of a general multi‐sensor information fusion architecture. An approach borrowed from sequential Bayesian processing is used to compute prior densities for joint Bayesian estimation‐detection. In this approach, a posteriori densities calculated at one sensor become priors at the next sensor after a coordinate transformation that transforms the outputs of each sensor to a common reference frame for all sensors. Reproducing prior densities are used to simplify the Bayesian iteration scheme and reduce communications requirements. The framework that is developed is equally applicable to networks where all sensors communicate with a centralized detection and estimation processor as well as those networks where sensors relay information from point to point. We anticipate using Bayesian iteration to convert posterior information into prior information on the next data gather cycle, iterat...