Evaluating large delay estimation techniques for assisted living environments

Abstract

Phase wraparound due to large inter-sensor spacings in multi-channel demixing limits the range of relative delays that many time–frequency relative delay estimators can estimate. The performance of a large relative delay estimation method, called the elevatogram, is evaluated in the presence of significant phase wraparound. This paper compares the elevatogram with the popular relative delay estimator used in DUET and the brute-force approach in D-AdRess and analyses its computational efficiency. The elevatogram can accurately estimate relative delays of speech signals of up to 800 samples, whereas DUET and D-AdRess were limited to delays of 7 and 35 samples, given a sampling rate of 16 kHz. Monte Carlo trials on 1000 real speech utterances show that the Matlab execution time of the elevatogram is slower than DUET, however it can accurately estimate delays that are 100-times greater than DUET. Source separation algorithms that use time–frequency relative delay estimators may be extended to function with maximum inter-sensor spacings of greater than 2.5 cm.