Effects of environmental, source, and monitoring parameters on the Degenerate Unmixing Estimation Technique algorithm in echoic environments

Blair R Conner, Purdue University

Abstract

This thesis researched the effects of environments, microphone spacing, and source variations on the blind sound source separation Degenerate Unmixing Estimation Technique (DUET). Sound analysis as a tool to monitor environments has become common in healthcare, consumer, and security industries. Monitoring sounds in an environment could also be applicable in the power industry to save electrical energy. The most common forms of sound analysis are sound identification and localization which use a variety of sound feature extraction and pattern recognition techniques. The most important and common problem for feature extraction in pattern recognition systems is the presence of noise from multiple sources. Sound source separation offers an innovative solution to improve sound identification and localization systems. The DUET algorithm is a blind sound source separation algorithm developed for anechoic mixtures, but previous research has suggested that the DUET algorithm's performance in echoic environments should be studied. This thesis aimed to determine if the DUET algorithm is capable of detecting and separating multiple unknown sound sources in an echoic environment. This research tested the DUET algorithm's performance for sound source detection and separation, to better understand the effects of the environment, microphone spacing, and source variations. The sample set consisted of two real and one clean environment, three microphone spacings, and fifteen sound source combinations. The data was collected using high quality professional grade recording equipment. Matlab was used to process the data and the researcher visually inspected all histograms and listened to all sound source estimations to analyze the DUET algorithm's performance. These results were finally assessed using a success rate percentage metric. Both the DUET algorithm's performance and the final results were analyzed for both physical and spatial sound sources. This research confirms that both sound source detection and separation of multiple unknown sound sources in echoic environments is possible using the DUET algorithm; however, the environment, microphone spacing, and source variations all have impacts, usually negative, on the DUET algorithm's performance in echoic environments. The overall success rates were on average around 50% for both source detection and separation when only two spatial sources were present. Experiments with three or four sources showed success rates below 25%. This suggests that the DUET algorithm, without modification, is not effective for practical source detection or separation in echoic environments. While many problems exist when using the DUET algorithm in echoic environments for source detection and separation, solutions and other applications are possible using the algorithm.

Degree

M.S.

Advisors

Panigrahi, Purdue University.

Subject Area

Electrical engineering

COinS