Analysis of a neural network approach to linear beamforming

Kuno, Y.M. University of Campinas, Brazil, yugo4k@decom.fee.unicamp.br

Masiero, B.S. University of Campinas, Brazil, masiero@unicamp.br

Madhu, N. Ghent University– imec, Belgium, nilesh.madhu@ugent.be

Abstract

Beamforming algorithms are applied to signals captured by sensor arrays to enhance signals arriving from de- sired directions while reducing background noise and other interfering signals. Beamforming coefficients can be optimally derived, usually satisfying some second-order statistical criterion such as the mean noise variance or mean output power, if the direction of arrival of the desired and interfering sources, as well as a statistical description of the background noise, is available. However, this is usually not the case. In an attempt to blindly estimate optimal beamforming coefficients directly from the array signals, we proposed the use of a neural net- work architecture. The network is composed of several blocks of neural network layers, each block responsible for processing one discrete frequency bin in the short-time discrete Fourier representation of the signals. The input to each block is the microphone signal vector at the corresponding discrete frequency, and for a single time frame. The output is the set of beamforming coefficients for that frequency and time frame. By analyzing the array's beampattern when the position of both the target source and the interferer varies we observed that the proposed neural network can track both sources, keeping the array pointed to the target source while nulling out the direction of the interferer. Thus it effectively serves as an adaptive beamformer. This paper evaluates this architecture rigorously based on several different performance metrics for the case of a single target plus localised or non-localised interference sources.

Keywords:

beamforming

neural network

blind source separation

speech enhancement.

References

[1] Michael Brandstein, Darren Ward. Microphone Arrays. Springer, 2001. ISBN 978-3-662-04619-7.

[2] Krim, H. e Viberg, M. Two decades of array signal processing research: the parametric approach. IEEESignalProcessing Magazine, 13(4):67–94, 1996. doi:10.1109/79.526899.

[3] Gannot, Sharon; Vincent, Emmanuel; Markovich Golan, Shmulik e Ozerov, Alexey. A consolidated perspective on multimicrophone speech enhancement and source separation. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 25(4):692 730, 2017. doi:10.1109/TASLP.2016.2647702.

[4] H. Moore, Alastair; Hafezi, Sina; R. Vos, Re becca; A. Naylor, Patrick e Brookes, Mike. A compact noise covariance matrix model for mvdr beamforming. IEEE/ACM Transactions on Audio, Speech, and Language Processing, págs. 1–14, 2022. doi:10.1109/TASLP.2022.3180671.

[5] Kuno, Yugo M.; Masiero, Bruno S. e Madhu, Nilesh. Aneural network approach to broadband beamform ing. Em Proc. 23rd International Congress on Acous tics, Aachen, Germany, 2019.

[6] UMA-8 v2 Product datasheet. miniDSP, 5 2018. v2.

[7] Yamagishi, Junichi; Veaux, Christophe e MacDonald, Kirsten. CSTR VCTK Corpus: English multi-speaker corpus for CSTR voice cloning toolkit (version 0.92), 2019.

[8] Rix, A.W.; Beerends, J.G.; Hollier, M.P. e Hek stra, A.P. Perceptual evaluation of speech qual ity (PESQ)-a new method for speech quality as sessment of telephone networks and codecs. Em 2001 IEEE International Conference on Acoustics, Speech, and Signal Processing. Proceedings (Cat. No.01CH37221), volume 2, págs. 749–752 vol.2, 2001. doi:10.1109/ICASSP.2001.941023.

[9] Taal, Cees H.; Hendriks, Richard C.; Heusdens, Richard e Jensen, Jesper. A short-time objective intel ligibility measure for time-frequency weighted noisy speech. Em 2010 IEEE International Conference on Acoustics, Speech and Signal Processing, págs. 4214 4217, 2010. doi:10.1109/ICASSP.2010.5495701.