Muhammad Z. Ikram

Permutation inconsistency in blind speech separation: investigation and solutions

Acoustic reverberation severely limits the performance of multiple microphone blind speech separation (BSS) methods. We show that the limited performance is due to random permutations of the unmixing filters over frequency. This problem, which we refer to as permutation inconsistency, becomes worse as the length of the room impulse response increases. We explore interesting connections between BSS and ideal beamforming, which leads us to propose a permutation alignment scheme based on microphone array directivity patterns. Given that the permutations are properly aligned, we show that the blind speech separation method outperforms the nonblind beamformer in a highly reverberant environment. Furthermore, we discover the tradeoff where permutations can be aligned by affording a loss in spectral resolution of the unmixing filters. We then propose a multistage algorithm, which aligns the unmixing filter permutations without sacrificing the spectral resolution. For our study, we perform experiments in both real and simulated environments and compare the results to the ideal performance benchmarks that we derive using prior knowledge of the mixing filters.

Non-linear acoustic echo cancellation using cascaded Kalman filtering

This paper presents a new paradigm of solving the non-linear acoustic echo cancellation problem. The non-linear echo path is modeled by a memoryless non-linearity followed by a linear FIR filter. The problem is cast into a state-space framework and solved using a cascade of Kalman filters in time domain, one filter adapting to the linear echo path and the other filter adapting to the memoryless non-linearity. It is shown that the proposed method outperforms the existing NLMS-based method in filter convergence and misalignment while enjoying an additional benefit of unsupervised and variable step-size control. Interesting connections will be made between the proposed method and the widely-known NLMS-based echo canceler. Practical recommendations are provided on implementing the proposed method efficiently on a general-purpose processor. Finally, simulation results are presented that exhibit its performance advantages.

Blind source separation and acoustic echo cancellation: A unified framework

We explore interesting connections between blind source separation (BSS) and acoustic echo cancellation (AEC), and develop a framework where the AEC problem is transformed and solved as a BSS problem. We show that by careful selection of the BSS algorithm the double-talk (DT) problem in AEC is solved without the need to use a DT detector or a step-size controller. Furthermore, the echo cancellation performance is maintained even during single-talk when only the far-end speaker is active. The algorithm converges to the true echo path much faster than the normalized least-mean squares adaptation. Moreover, the proposed algorithm does not require a knowledge of the echo-tail length and is robust against under estimation of the echo-filter length. The simple implementation and fast convergence of the proposed method makes it a suitable candidate for implementation on low-power general purpose DSPs.

Implementation of blind source separation and a post-processing algorithm for noise suppression in cell-phone applications

For cell-phone applications, single microphone noise suppression techniques have limited performance at very low SNR (close to 0 dB). In certain cases, they also suffer from the artifacts of nonlinear processing. In this paper, we will show that techniques based on two-microphone blind source separation (BSS) algorithm provide significant interference suppression for cell-phone applications, particularly at low operating SNR values. We also propose a post-BSS processing method based on frequency-domain spectral subtraction that further improves the BSS speech output in diffused noise case. We optimize the BSS algorithm so that it can be implemented on a low-power audio codec processor, which would be ideal for cell phone applications. Furthermore, based on extensive analysis under different noise and acoustic conditions, we suggest recommendations for optimal placement of microphones on a cell phone. We also study the trade-off between the unmixing filter length and the noise suppression performance. In all our experiments, we use real recordings made on a cell phone equipped with two microphones.

Double-talk detection in acoustic echo cancellers using zero-crossings rate

We propose a new method to detect double-talk and control filter adaptation in an acoustic echo canceller (AEC). The method is based on computing the zero-crossings rate (ZCR) of the AEC output and comparing it against a suitably-chosen threshold. As the ZCR values falls below the threshold, double talk is declared and the AEC filter adaptation is either slowed down or halted. The zero crossings are very easy to compute by observing the sign changes of two consecutive samples from the output of the AEC. In contrast to most existing methods, the computational burden of the proposed method is minimal and it can, therefore, be conveniently implemented on a low-power, low-resource processor. This computational simplicity is enjoyed without sacrificing for any AEC performance. We will illustrate effectiveness of the proposed method by comparing against the existing state of the art and present guidelines on choosing parameters for computing the sample-by-sample ZCR.

3-D object tracking in millimeter-wave radar for advanced driver assistance systems

This paper presents a new 3-D tracking method for millimeter-wave radar imaging of objects around a car in advanced driving assistance systems (ADAS). Current automotive radar systems only estimate objects in 2-D plane (range and azimuth) and hence only a 2-D tracking is employed. We expect future evolutions of these systems to provide the additional elevation information allowing a 3-D view around the car. Given the estimated locations of multiple objects, the tracking algorithm refines the estimates and initiates new tracks as new objects are detected in the scene. The tracking is based on Extended Kalman filter that takes measurements in spherical coordinates and relates them to state vector in Cartesian coordinates. Furthermore, estimates of object radial velocity are related to object velocities in Cartesian coordinates. We will present simulation results to illustrate the effectiveness of the proposed tracking model.

Joint antenna-array calibration and direction of arrival estimation for automotive radars

We present an iterative method for joint antenna-array calibration and direction of arrival estimation using millimeter-wave (mm-Wave) radar operating at 77 GHz. The calibration compensates for antenna-array coupling, and phase and gain errors, and does not require any training data. This method is well suited for applications, such as automotive radars, where multiple antenna elements are packaged on a chip and where offline calibration is either expensive or is not possible. The proposed method is highly effective when the array coupling is a function of direction of arriving waves from the object. The novel optimization method proposed in the paper allows it to be used for a 2D array of any shape. Experiments using real data collected from a four-element array on a single-chip radar demonstrate the viability of the algorithm.

A new data association method for 3-D object tracking in automotive applications

We present a new method for data association in 3-D object tracking for automotive applications. The method is a variant of the nearest-neighbor data association and is based on comparing the location of an existing track with that of each incoming object and associating to the one which is closest in 3-D space. As a pair is associated, it is removed from the search space and the association process continues until all assignments are made. Our experiments show that the proposed method significantly reduces the processing cost as compared to the existing full-search nearest-neighbor method and maintains similar performance at the signal to noise ratios that are typically encountered in automotive object tracking. We will provide guidelines on selecting the operating parameters and suggestions on handling the case when the number of incoming objects is not equal to the number of existing tracks.

Promoting convergence in multi-channel blind signal separation using PNLMS

The proportionate normalized least-mean squares (PNLMS) adaptation algorithm exploits the sparse nature of acoustic impulse responses and assigns adaptation gain proportional to the absolute value of filter coefficients, thereby resulting in faster convergence. In the past it has shown to improve convergence of acoustic paths in echo-cancellation applications. In this paper, we investigate the use of PNLMS algorithm for blind speech separation and show that with a careful selection of operating parameters the PNLMS algorithm greatly helps promote convergence of the un-mixing filters when compared to the conventional normalized least-mean-squares (NLMS) adaptation. The PNLMS based blind speech separation is suitable for real-time implementation as it promises faster convergence and requires only a modest increase in complexity as compared to the NLMS algorithm.

On structured total least squares for blind identification of multichannel FIR filters

Structured total least-squares (STLS) provides a nice framework for approximating a full-rank affmely-structured matrix with a rank-deficient matrix having the same affine structure. In this paper, we investigate the use of STLS method for blind identification of multiple FIR channels driven by an unknown deterministic input. First, we exploit the block - Hankel affine structure of the data matrix, which motivates the use of STLS-based methods. Then, we derive an iterative non-linear solution to the unknown channel parameters by using a generalized form of singular value decomposition. We carry out extensive numerical simulations to compare the performance of the proposed method against the well-known least-squares (LS) method, where the affine structure of the date matrix is overlooked. These results reveal that the STLS based method outperforms the LS method for ill-conditioned as well as well-conditioned channels over a wide range of SNR.