Adaptive Algorithms for the Identification of Time-Variant Acoustic Systems

Authors:

Kühl, S.

Ph. D. Dissertation

 

School:

IKS, RWTH Aachen University

Adress:

Templergraben 55, 52056 Aachen

Series:

Aachen Series on Communication Systems

Number:

1

Date:

2022

ISBN:

978-3-84408-633-1

DOI:

10.18154/RWTH-2022-07587

Language:

English

Abstract

Many digital speech and audio communication systems incorporate models of acoustic systems during signal processing. Often, the impulse responses (IRs) of these acoustic systems have to be identified during or before using a communication system by means of adaptive algorithms.
Acoustic systems describe how sound is affected during transmission between a source and a receiver. Examples of acoustic systems comprise rooms with reflections from boundaries and scattering from objects, communication devices such as smartphones or smart home devices, or even a human head where shadowing effects occur. In many situations, the IRs of these systems need to be identified. Possible scenarios are acoustic measurements or system identification in speech communication applications, e.g., for acoustic echo cancellation (AEC). Depending on the specific use case, certain aspects have to be taken into account. For a measurement the excitation signal can be designed, whereas for speech communication applications the system identification relies on the communication signal. Hence, for the latter case, correlation has to be considered during the system identification. In addition, the systems to be identified may vary over time when the acoustic environment changes, e.g., due to moving objects, air movements, or temperature changes. Therefore, adaptive algorithms must be used to track the system’s state. Additional challenges arise when considering the identification of multiple channels simultaneously.
This thesis considers the different aspects of system identification of time-variant acoustic systems for diverse scenarios. It provides new insights into the relationship of different system identification algorithms and proposes novel concepts and algorithms for specific applications. %Below, the major contributions are described.
Measurement procedures from the field of acoustics and tracking algorithms from communication applications are compared in a joint mathematical framework, which contributes a novel proof of their mathematical equivalence for periodic excitation, thus helping to bridge the gap between these two closely related fields. It is shown that the excitation signal mainly determines the identification behavior. Whereas for time-invariant scenarios the achievable quality of the measurement is often assessed in terms of signal-to-noise ratio (SNR), in the time-variant case the tracking behavior over time is also very important. Therefore, different excitation signals are developed. Sweep signals are often used for measurements. For tracking algorithms, noise signals are considered to achieve good tracking performance. This thesis provides insights into the adaptation behavior over time in dependence on the excitation signal used and presents a new multi-channel excitation strategy for continuous system identification based on exponential sweeps (ESs).
For AEC in hands-free calls or teleconferences, acoustic systems have to be tracked with correlated excitation. Many different algorithms have been developed for this purpose, e.g., the normalized least mean square (NLMS) algorithm, the recursive least squares (RLS) algorithm, and the Kalman filter. In this thesis, it is shown how these algorithms can be derived as constrained versions of a general Kalman filter. This shows the close relationship between theses algorithms and provides new insights into their mode of operation. For the time-domain Kalman filter, a novel convergence analysis is carried out, showing how correlated excitation affects the adaptation performance. For the frequency-domain Kalman filter, a decorrelation stage is developed, which improves the adaptation performance for correlated excitation. Several estimation methods for a priori information employed in the Kalman filter are presented and the dependency on the specific adaptation algorithm used is shown.
Furthermore, acquiring multiple channels of an acoustic system simultaneously in order to capture the full sound field offers the potential to provide a more natural sound experience. When dealing with multiple channels, additional aspects and difficulties compared to identifying only one channel arise, which are analyzed in this thesis. The non uniqueness problem (NUP) concomitant with multiple correlated channels is analyzed and a novel decorrelation approach based on techniques from primary ambient extraction (PAE) is proposed in this thesis. Moreover, a reduced-complexity algorithm for spatial audio communication is presented relying on a sparse activity of sound sources.
For many applications, besides the input and output signal, further information is available which can be incorporated in the filter adaptation process as a priori information. Examples of a priori information in this context may include additional signals, system states, and general information about the application at hand that the identification algorithm can profit from. This thesis proposes methods to use a priori information for the combination of a beamformer (BF) and AEC and for adaptive feedback cancellation (AFC) as required for IP based teleconferencing.

Download

BibTeX

Copyright © by IKS
kuehl22.pdf

This material is presented to ensure timely dissemination of scholarly and technical work. Copyright and all rights therein are retained by authors or by other copyright holders. All persons copying this information are expected to adhere to the terms and constraints invoked by each author's copyright. In most cases, these works may not be reposted without the explicit permission of the copyright holder.

---