Kensaku Fujii, Fujitsu Laboratories Ltd (Japan)
Juro Ohga, Fujitsu Laboratories Ltd (Japan)
This paper presents a method of differentiating between double-talk and echo path change requisite for holding the acoustic coupling gain stable. In an acoustic echo canceller system, powers of the reference signal and the environmental noise incessantly fluctuate. The conventional system has been designed so as to suspend the estimation process while either power is outside a desirable range. This paper introduces a method to continue the estimation process even in that while, by using the block implementation of the normalized least mean square (NLMS) algorithm and by adjusting that block length.
Athanasios Liavas, Institut National des Télécommunications (France)
Phillip Regalia, Institut National des Télécommunications (France)
The adequateness of IIR models for acoustic echo cancellation is a long standing question and the answers found in the literature are conflicting. We use results from rational Hankel norm and least-squares approximation and we recall a test which provides a priori performance levels for FIR and IIR models. We apply this test to measured acoustic impulse responses. Upon comparing the performance levels of equal complexity FIR and IIR models, we do not observe any significant gain from the use of IIR models. We attribute this phenomenon to the shape of the energy spectra of the acoustic impulse responses, so tested, which possess many strong and sharp peaks. Faithful modelling of these peaks requires many parameters irrespective of the type of the model.
Peter Hansen, Technical University of Denmark (Denmark)
Per Christian Hansen, Technical University of Denmark (Denmark)
Steffen Duus Hansen, Technical University of Denmark (Denmark)
John Sorensen, Technical University of Denmark (Denmark)
In this paper the signal subspace approach for nonparametric speech enhancement is considered. Traditionally, the SVD (or the eigendecomposition) is used in frame-based methods to decompose the vector space of the noisy signal into a signal-and noise subspace [1, 2, 51. Linear estimation of the clean signal from the information in the signal subspace is then performed using a set of nonparametric estimation criteria. In this paper, the rank-revealing ULV decomposition is used instead of the SVD, and we use recursive updating of the estimate instead of working in frames. An ULV formulation of three different estimation strategies is considered: Least Squares, Minimum Variance and Time Domain Constrained. Experiments indicate that the ULV-based algorithm is able to achieve the same quality of the reconstructed speech signal as the SVD-based method.
Nenad Bojic, Victoria University of Technology (Australia)
Developing low-cost Acoustic Echo Cancellation (AEC) systems can be a difficult task due to the hunted internal memory available on low cost Digital Signal Processors (DSP). This paper introduces the Partial Room Characteristic Cloning (PRCC) AEC that has an inherently low memory requirement and computational rate. Real-time experimental results indicate that the PRCC AEC can achieve an ERLE of 15 - 20 dB with 200 words of memory and a computational rate of 0.8 MIPS, making it ideal for low cost AEC applications.
Joerg Meyer, University of Bremen (Germany)
Klaus Uwe Simmer, University of Bremen (Germany)
Karl Dirk Kammeyer, University of Bremen (Germany)
In this paper, we compare several one-channel and two-channel noise-estimation techniques. We focus on their estimation features in a non-stationary noisy environment while a speech signal is present, as this is one of the unsolved problems for spectral subtraction algorithms. All one-channel solutions make use of the different statistics of speech and unwanted noise. The two-channel algorithms use the spatial characteristics of the noise field in order to estimate the power spectral densities (PSD). First, we will briefly describe several existing algorithms, then we will introduce a new one which is related to the one proposed by Gierl [11.
August Kaelin, Swiss Federal Institute of Technology (ETH) (Switzerland)
Sigisbert Wyrsch, Swiss Federal Institute of Technology (ETH) (Switzerland)
This paper describes a digital hearing aid realized in the frequency domain that compensates for recruitment of loudness and cancels acoustic echos. In contrast to conventional systems which are based on a noise-probe signal, our echo canceller is adapted using only the available (e.g. speech) input signal. The main problems caused by a nonlinear feedforward filter, for compensating recruitment of loudness, are discussed using analytical results of the steady state behavior of the closed-loop hearing-aid system. The proposed solutions have been implemented and tested on a dummy behind-the-ear (bte) hearing-aid device.
Joe Timoney, Trinity College (Ireland)
Brian Foley, Trinity College (Ireland)
The LMS-driven Adaptive Periodic Noise Canceller (APNC) can be used for acoustic echo suppression in the hands-free situation. Only recently has it been determined that the LMS algorithm can be robust [11. The purpose of this contribution is to apply the concept of robustness, as derived from [11, to the APNC, thereby introducing a useful quantitative measure of performance. This is achieved through an analysis of the performance of the APNC in an open-loop feedback system. Through the application of H-theory, conditions are shown under which the APNC, driven by the LMS algorithm, will exhibit robust performance properties. This has a direct application to the use of the APNC in echo control applications.
Martin Medvecky, Slovak University of Technology
S. Herrera-Garcia, Research Center - Cinvestav (Mexico)
The implementation of modified version Of Normalised Least-Mean-Square algorithm (M-NLMS) will be presented. The M-NLMS algorithm was developed for efficient weight adaptation of acoustic echo canceller based on modified finite impulse response (M-FIR) filter . The M-FIR filter is especially suitable for hardware realisation of acoustic echo canceller. The M-NLMS algorithm enables maximum exploitation of advantages of the M-FIR filter. The proposed adaptive algorithm modification and implementation principles can be applied to other LMS based algorithms as well. The comparison of acoustic echo cancellers with adaptive FIR filter adapted by NLMS algorithm d adaptive M-FIR filter adapted by M-NLMS algorithm are presented.
Frank Heinle, Universität Erlangen-Nürnberg (Germany)
R. Rabenstein, Universität Erlangen-Nürnberg (Germany)
A. Stenger, Universität Erlangen-Nürnberg (Germany)
Competitive audio consumer products require not only cheap signal processing hardware but also low-cost analog equipment and sound transducers. The nonlinear distortions produced by these electroacoustic transmission systems cannot be described and analyzed by standard methods based on linear systems theory alone. In order to take the nonlinear properties into account, we present a measurement method for the linear and nonlinear transmission characteristics of almost arbitrary systems and show its application to the analysis of electro-acoustic systems. Examples demonstrate the measurement Of the impulse response of a loudspeaker-enclosure microphone-system with cheap analog equipment.
Tilman Huhn, Zentrum Mikroelektronik Dresden (Germany)
Hans-Joachim Jentschel, Technische Universität Dresden (Germany)
This paper refers to the multi-channel system identification of a linear single-input multiple-output (SIMO) system by an adaptive filter based on the normalised LMS algorithm. Configurations with an SIMO echocanceler are of special interest for the combination of echo cancellation and noise reduction principles. In this paper we investigate the problem of optimising a generalised adaptation parameter, called the stepsize matrix. The solution leads to a novel adaptation control approach, realising a coupling between the different channels. We will show that the channel coupling allows a partial cancellation of the measurement noise, which is adverse to the identification accuracy.
Christina Breining, Darmstadt University of Technology (Germany)
In this paper, we derive a new optimum time-variant stepsize for the adaptation of an echo cancellation filter. The optimum stepsize is then used in the form of cost functions to evaluate the performance of stepsize control methods. Finally, we show first results of combining a number of estimators for stepsize control by means of a neural network trained on this stepsize and a suitable cost function.
Bernhard Nitsch, Technische Hochschule Darmstadt (Germany)
Acoustic echo cancellers for hands free telephony require filter impulse response sizes between 128 and 256 ms to reach a sufficient error return loss enhancement. Using a sampling rate of 16 kHz yields filter sizes of 2000 to 4000 coefficients. For real time implementations on digital signal processors (DSPs) time domain based adaptive filters need huge processing power. Filtering in frequency subbands [51 or using block adaptive filter algorithms can reduce the algorithm complexity significantly. In this paper the Partitioned Exact Frequency Domain Block NLMS (PEF13NLMS) algorithm is presented which is mathematically an exact formulation of the time domain NLMS algorithm. The PEFBNLMS algorithm combines a computational complexity reduction of 3 to 5 times compared to the NLMS algorithm with the same tracking ability.
Thomas Schertler, Darmstadt University of Technology (Germany)
Gerhard Uwe Schmidt, Darmstadt University of Technology (Germany)
Hands-free telephone sets with adaptive echo cancellers need an adaptive loss control in order to guarantee a loss of acoustic echo of about 45 dB (ITU-T Recommendation G.167 [11). The quality of the speech transmission depends on the coupling between the echo canceller and the loss control. In this contribution we present an implementation of a hands-free telephone algorithm, using a low-cost fixed-point DSP which is well suited for integration in consumer devices.
Valerie Turbin, CNET DIH/CMC (France)
André Gilloire, CNET DIH/CMC (France)
Pascal Scalart, CNET DIH/CMC (France)
Christophe Beaugeant, CNET DIH/CMC (France)
We propose a general approach based on optimal filtering and use of psychoacoustic constraints to achieve acoustic echo cancellation which is applied in two contexts: teleconferencing and mobile telephones in cars. In the teleconferencing context, the acoustic echo cancellation system is composed of a conventional echo canceller combined with an optimal filter. In the mobile telephony context where not only the acoustic echo but also the ambient noise are to be cancelled, we propose to reduce globally both disturbances with only one optimal filter. We show that using psychoacoustic criteria in the optimal filter computation enables to reduce the distortion generated on the near-end speech especially when the perturbator is the acoustic echo.
Christiane Antweiler, Aachen University of Technology (Germany)
Adaptive filters are used extensively in telecommunication e.g. for acoustic echo control, noise reduction or channel equalization. The so called normalized least mean square (NLMS) algorithm represents one of the most widely used gradient-based adaptation algorithm. One of the main problems of the NLMS algorithm in an echo control application is that the adaptation is driven by speech signals, i.e. spectrally colored signals, which reduce the convergence speed significantly. In this paper an alternative approach to several existing algorithms (e.g. [1,6,7]) is introduced, which is based on a priori knowledge of the spectral characteristics of the room impulse response in form of spectral weighting filters. This technique results in an accelerated convergence rate and benefits from a more robust behavior in the presence of background noise, such as occurring for example in a car. Besides the theoretical derivation of the proposed algorithm its realization aspects, complexity features, and simulation results are discussed in detail.
Karim Maouche, Institut EURECOM (France)
Dirk Slock, Institut EURECOM (France)
In this paper, we derive a new normalisation technique for Frequency Domain Adaptive Filtering (FDAF) algorithms. FDAF algorithms are well-known for their low computational complexity and for their decorrelating property that allows the use of different step sizes for each adaptive weight, yielding a uniform convergence of all the modes of the input signal. In these algorithms, normalisation is done by recursively estimating the power of each frequency bin. By introducing a normalisation based on an orthogonal projection (as is the case for the Affine Projection algorithm), we derive new frequency domain adaptive filtering algorithms and show by means of simulation that the convergence speed is improved.