Home Home
Andreas Englesberg, Kiel University (Germany)
Thomas GuIzow, Kiel University (Germany)
Spectral subtraction is a popular method for speech enhancement, if the speech signal is corrupted by additive noise. It is based on the manipulation of the magnitude of the noisy speech spectrum. Previous realizations used linearly spaced frequency transformations. We propose the application of two filterbanks with bark-scaled frequency bands: a discrete wavelet transform and a nonuniform polyphase filterbank. The enhancement results are compared to those obtained with uniform spectral transformations.
ae970028.pdf (Scanned)
TOPDariusz Bismor, Silesian Technical University (Poland)
The goal of the paper is to present results on research on applying different algorithms and structures to active noise control in an acoustic duct. A few modifications of Least Mean Squares (LMS) algorithm are presented and compared. Two different variations of feedforward control structure has been considered.
ae970033.pdf (Scanned) TOP
Q.G. Liu, Nortel Technology (Canada)
B. Champagne, INRS - Télécommunications (Canada)
Subband adaptive filtering is an important application of filter banks in which maximally decimated filter banks can not be used in general because of decimation aliasing effects. This leads to the use of oversampling schemes in the filter bank design wherein the perfect reconstruction (PR) or near PR property is still required. In this work, a simple design technique for uniform DFT filter bank with near PR property is presented for the purpose of subband adaptive filtering. The prototype filter in the proposed filter banks can be obtained simply by performing an interpolation of a two-channel QMF filter. The filter bank design technique presented in this paper is of particular interest in engineering applications, as demonstrated by design examples.
ae970035.pdf (Scanned) TOP
Moritz Harteneck, University of Strathclyde (U.K.)
R.W. Stewart, University of Strathclyde (U.K.)
J.M. Paez-Borrallo, Ciudad Universitaria (Spain)
Recently a new real-valued oversampled filter bank has been proposed which reduces the "inband" alias. The filterbank consists of at least three channels which are subsampled by different subsampling ratios. In this paper we investigate into the applicability of this filterbank for adaptive subband filtering and compare the setup with existing subband and fullband techniques.
ae970038.pdf (Scanned) TOP
Akira Nakagawa, NTT Human Interface Laboratories (Japan)
Yoichi Haneda, NTT Human Interface Laboratories (Japan)
Shoji Makino, NTT Human Interface Laboratories (Japan)
This paper presents a subband acoustic echo canceller (SBEC) using two different analysis filters and an 8th order complex affine projection algorithm (APA). This SBEC uses different analysis filters to divide the input signals and echo signals. The analysis filter for the input signals is designed to improve the convergence speed, and the analysis filter for the echo signals is designed to achieve high echo return loss enhancement. This SBEC also uses an 8th order complex APA to further improve the convergence. We implemented the proposed SBEC on digital signal processors and evaluated it. The results show that our SBEC converges five times faster than the conventional SBEC.
ae970039.pdf (Scanned) TOP
Oguz Tanrikulu, Imperial College of Science Technology and Medicine (U.K.)
Anthony Constantinides, Imperial College of Science Technology and Medicine (U.K.)
For highly selective filter-banks the aliasing in subband acoustic echo cancellers (AEC) is confined into very narrow spectral regions. Therefore, narrowband notch filters were introduced into the analysis-banks in order to attenuate the aliasing prior to echo cancellation. Two contributions are presented in this paper. Firstly, it is shown that the notch filtering operation can be implicitly performed by using Cascaded Power Symmetric IIR (PS-IIR) filter-banks. Secondly, the adaptive algorithms running in neighbouring subbands must be coupled via continuity constraints. Therefore the well-known NLMS algorithm is modified and the Continuity Constrained NLMS (CC-NLMS) algorithm is proposed. The simulation results confirm the attenuation of the aliased components.
ae970041.pdf (Scanned) TOP
Stephan Weiss, University of Strathclyde (U.K.)
Moritz Harteneck, University of Strathclyde (U.K.)
R.W. Stewart, University of Strathclyde (U.K.)
Adaptation of the tap profile in subband adaptive system identification problems can further enhance the efficient use of computational resources if implemented on a DSP with an otherwise too tight benchmark performance. Here, we derive a generalization of previous work to extend tap-assignment algorithms to a new class of oversampled filter banks with non-uniform bandwidths and different subsampling ratios. We compare efficiency and adaptation results for this approach to the critically sampled case and a fullband identification with same complexity.
ae970043.pdf (Scanned) TOP
Sumitaka Sakauchi, NTT Human Interface Laboratories (Japan)
Yoichi Haneda, NTT Human Interface Laboratories (Japan)
Shoji Makino, NTT Human Interface Laboratories (Japan)
The frequency characteristic of the echo return loss requirement (ERLRf) was investigated using subjective assessments. The ERLRf is an important factor in the design and performance evaluation of a subband echo canceller (SBEC). The ERLRf during single-talk was obtained as attenuated band-limited echo levels that subjects did not find objectionable when listening to test speech and its band-limited echo under various transmission conditions. When we investigated the ERLRf during double-talk, subjects also heard nearend speech. Here, the echo was limited to a 250-Hz bandwidth assuming the use of an SBEC with 32 subbands. The test results showed that: (1) as the transmission delay rose above 100 ms, the ERLRf at lowfrequency bands around 1 kHz increased significantly; (2) when the room reverberation time is relatively long (about 450ms), the ERLRf increases especially at lowfrequency bands around 1 kHz even if the transmission delay is short (28 ms); and (3) the ERLRf during double-talk is about 5 to 10 dB lower than during single-talk. The obtained ERLRf will be useful for designing an efficient SBEC.
ae970044.pdf (Scanned) TOP
Peter Eneroth, Lund University (Sweden)
Tomas Gansler, Lund University (Sweden)
In the delayless subband adaptive echo canceller, the estimated subband impulse responses are non-causal and the non-causal parts are truncated. In the closedloop structure, the feed-back of the error-signal compensate for the truncation, which is not the case in the open-loop structure. A modified open-loop structure is proposed, in which the truncation error is reduced. The performance of the proposed structure approaches the closed-loop structure, and has the advantage of a higher convergence rate.
ae970047.pdf (Scanned) TOP
Stefan Gustafsson, Aachen University of Technology (Germany)
Rainer Martin, Aachen University of Technology (Germany)
In this paper an acoustic echo compensator with an additional frequency domain adaptive filter for combined residual echo and noise reduction is proposed. The algorithm delivers high echo attenuation as well as high near end speech quality over a wide range of signal-to-noise conditions. The system makes use of a standard time domain echo compensator of low order, after which the proposed adaptive filter is placed in the sending path. In contrast to other combined systems [1, 2, 3], our method uses an explicit estimate of the power spectral density of the residual echo after echo compensation. The separate estimations of the power spectral densities of the residual echo and of the background noise, respectively, are then flexibly combined, such that in the processed signal a low level of intentionally left background noise will effectively mask the residual echo.
ae970048.pdf (Scanned) TOP
Desmond Phillips, Loughborough University (U.K.)
Colin Cowan, Queen's University of Belfast (U.K.)
Each subband in a multirate acoustic echo canceller has different statistical properties. As there is a wide range of adaptive filters to choose with different cost 1 performance trade-offs, it is important to target computational resources intelligently depending on the perceptual contribution of each subband in the overall error residual. This results in the scenario of a heterogeneous AEC. For this optimisation problem, both a pertinent methodology and reliable source data through comprehensive benchmarking are required. This paper discusses some of the methodological issues raised by AF benchmarking in subbands.
ae970055.pdf (Scanned) TOP
Beghdad Ayad, Université de Rennes (France)
Regine Le Bouquin-Jeannes, Université de Rennes (France)
Acoustic echo and noise cancelling is fundamental in any speech transmission system. In the solutions addressed to this problem, the acoustic echo cancellation is carried out by identification of the transfer function of the acoustic channel. In this paper, another approach is proposed where echo cancellation is realized by filtering the microphone observation. Within this approach, three systems are developed. For noise reduction, an updating of the noise characteristics in the presence of speech is studied. Measures of echo return loss enhancement, noise reduction and speech distortion are presented. It happens that the new approach performs better than the basic one.
ae970058.pdf (Scanned) TOP
Amir Hussain, University of Paisley Scotland (U.K.)
Douglas R. Campbell, University of Paisley Scotland (U.K.)
A general class of single-hidden layered, linear-in-the-parameters feedforward Artificial Neural Networks is proposed for processing band-limited signals in a multimicrophone sub-band adaptive speech enhancement scheme. The sub-band spacing within the adaptive speech enhancement system is set according to a published cochlear function. Comparative results achieved in simulation experiments demonstrate that the proposed sub-band scheme is capable of significantly outperforming conventional full-band and sub-band noise cancellation methods employing linear processing, in the presence of non-linear interference.
ae970059.pdf (Scanned) TOP
Rhonda Wilson, Meridian Audio Limited (U.K.)
Patrick Naylor, Imperial College (U.K.)
Mike Brookes, Imperial College (U.K.)
This paper investigates performance limitations in subband acoustic echo controllers due to modelling errors. It is shown that the subband model of a system is exact only when the number of taps in each subband filter is infinite. When the number of taps used is finite, it is shown that the system modelling error represents an upper bound on performance.
ae970073.pdf (Scanned) TOP
Pia Dreiseitel, unknown
Henning Puder, unknown
(No abstract available for this paper)
TOP