부산대학교 도서관

Acoustic sensor networks with ad-hoc topology currently seek application for acoustic signal processing in place of traditionally compact microphone arrays. Wireless configuration of the network will, however, impose additional challenges of which the synchronization of autonomous clocks has already received considerable attention under the premise of lossless transmissions. Further constraining to low latency, a serious impact of packet loss is expected as a result of UDP protocol operating the network. Such packet loss will be detrimental not only to the function of multi-channel acoustic signal processing over the network, but may deteriorate already the necessary estimation of sampling rate offset (SRO) between ad-hoc sensor nodes from incomplete waveforms. To study this aspect, the paper firstly revisits formal statistical packet loss models based on first-order Markov chains and, secondly, introduces online-capable methods for blind SRO estimation including the recursive band-limited interpolation (RBI), an online form of weighted averaged coherence drift (WACD) and a double-cross-correlation processor (DXCP) in FFT-domain. For experimental investigation, we then embed statistical simulation of burst-packet loss in a range of packet-loss rates into otherwise real acoustic data. In this setting, WACD and DXCP methods with proposed modifications demonstrate effectiveness in terms of SRO estimation accuracy and synchronization performance.