T. van Waterschoot

Analytical Expressions for the Power Spectral Density of CP-OFDM and ZP-OFDM Signals

In this letter, analytical expressions are derived for the power spectral density (PSD) of orthogonal frequency division multiplex (OFDM) signals employing a cyclic prefix (CP-OFDM) or zero padding (ZP-OFDM) time guard interval. Under the relatively weak assumptions that (i) the data are independent and identically distributed on all OFDM subcarriers and (ii) the OFDM pulse shape is sufficiently localized in time, simple closed-form PSD expressions can be obtained. These expressions are then compared to existing OFDM PSD expressions and validated by inspecting the power spectra of some standardized OFDM signals.

Double-Talk-Robust Prediction Error Identification Algorithms for Acoustic Echo Cancellation

The performance of an acoustic echo canceller may be severely degraded by the presence of a near-end signal. In such a double-talk situation, the variance of the echo path estimate typically increases, resulting in slow convergence or even divergence of the adaptive filter. This problem is usually tackled by equipping the echo canceller with a double-talk detector that freezes adaptation during near-end activity. Nevertheless, there is a need for more robust adaptive algorithms since the adaptive filters convergence may be affected considerably in the time interval needed to detect double-talk. Moreover, in some applications, near-end noise may be continuously present and then the use of a double-talk detector becomes futile. Robustness to double-talk may be established by taking into account the near-end signal characteristics, which are, however, unknown and time varying. In this paper, we show how concurrent estimation of the echo path and an autoregressive near-end signal model can be performed using prediction error (PE) identification techniques. We develop a general recursive prediction error (RPE) identification algorithm and compare it to three existing algorithms from adaptive feedback cancellation. The potential benefit of the algorithms in a double-talk situation is illustrated by means of computer simulations. It appears that especially in the stochastic gradient case a huge improvement in convergence behavior can be obtained

A Pole-Zero Placement Technique for Designing Second-Order IIR Parametric Equalizer Filters

A new procedure is presented for designing second-order parametric equalizer filters. In contrast to the traditional approach, in which the design is based on a bilinear transform of an analog filter, the presented procedure allows for designing the filter directly in the digital domain. A rather intuitive technique known as pole-zero placement, is treated here in a quantitative way. It is shown that by making some meaningful approximations, a set of relatively simple design equations can be obtained. Design examples of both notch and resonance filters are included to illustrate the performance of the proposed method and to compare with state-of-the-art solutions.

Adaptive feedback cancellation for audio signals using a warped all-pole near-end signal model

Sound amplification systems having a closed signal loop often suffer from acoustic feedback, which limits the achievable amount of amplification and severely affects sound quality. A promising solution to the feedback problem consists in predicting the feedback signal using an adaptive filter, however, a bias is then introduced due to signal correlation. In speech applications, a prediction-error-method- based approach to adaptive feedback cancellation has proven to be capable of providing sufficient decorrelation without sacrificing speech quality. This approach, which is based on estimating an all-pole near-end signal model, appears to be unappropriate for musical audio signals because of their large degree of tonality. We propose a novel prediction-error-method-based adaptive feedback cancellation algorithm that features a frequency-warped all-pole near-end signal model, which is better suited for tonal audio signals. Simulation results show a doubling of the convergence speed, with only a relatively small increase in computational complexity.

Partitioned block frequency domain Kalman filter for multi-channel linear prediction based blind speech dereverberation

The multi-channel linear prediction framework for blind speech dereverberation has gained increased popularity over the recent years. While adaptive dereverberation is desirable, most multichannel linear prediction algorithms are based on either batch or iterative frame-by-frame processing, where individual frames are treated independently. In this paper, we derive a partitioned block frequency domain Kalman filter that offers adaptive processing. The so-called excessive whitening problem is avoided by including an estimate of the target speech signal coloration in the filter update. The impact of constraining the state covariance matrix is discussed. The convergence behavior of the algorithm is evaluated in terms of the evolution of the room acoustical parameters direct-to-reverberant ratio, clarity index and early decay time, indicating good dereverberation performance.

A quantitative comparison of blind C 50 estimators

The problem of blind estimation of the room acoustic clarity index C50 from single-channel reverberant speech signals is presented in this paper. We analyze the performance of several machine learning methods for a regression task using 309 features derived from the speech signal and modeled with a Deep Belief Network (DBN), Classification And Regression Tree (CART) and Linear Regression (LR). These techniques are evaluated on a large test database (86 hours) that includes babble noise and reverberation using both artificial and real room impulses responses (RIRs). All methods are trained on a database which contains noise, speech and simulated RIRs different from the test set. The performance results show that the DBN model gives the lowest error for the simulated RIRs whereas the LR model gives the best generalization performance with the highest accuracy for real RIRs.

Speech dereverberation by data-dependent beamforming with signal pre-whitening

Among different microphone array processing techniques, data-dependent beamforming has been proven to be effective in suppressing ambient noise. When applied for dereverberation, however, the adaptation process results in a biased estimate of the beamformer coefficients leading to strong distortions at the beamformer output. In this paper, we investigate the origin of this bias for the generalized sidelobe canceller. It is shown that an unbiased estimate of the beam-former coefficients and thus dereverberation can be achieved if the source signal is a white random signal. Based on these findings, a pre-whitening approach for speech signals is proposed and combined with a generalized sidelobe canceller for speech dereverberation. The concept is demonstrated for the case of stationary speech-shaped noise as a source signal.