Improved angular resolution for sound source localization in a reverberant environment
Embargo Date
2021-05-07
OA Version
Citation
Abstract
Sound source localization is the problem of locating a speaker in three dimensions using data collected by microphones. It has a wide array of applications including video conferencing to steer a camera, robotic navigation when visual data fails, and as an intermediate step in sound source separation and noise reduction. In a reverberant environment, such as a conference room, the signal received at each microphone is the sum of the direct path of the signal from the speaker to the microphone and the signal reflecting off of walls and other surfaces in the room before reaching the microphones. Each of these reflected components can be modeled as additional sources. Effective sound source localization algorithms must be robust to these additional sources. We propose applying a bandpass filter to only consider higher frequencies to increase the angular resolution and improve the elevation estimate of a time difference of arrival based maximum-likelihood algorithm for sound source localization.
Results show that applying a bandpass filter improves angular resolution so that the elevation is more accurately measured. The average elevation error is reduced by as much as 90% with the proposed improvement. This improvement enables the use of smaller microphone arrays, even though they have worse angular resolution. Both simulated data and data collected in real conference rooms are used to substantiate the results.