Parameter selection for segregating speech from background noise
Loading...
Date
2015-05
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
The Ohio State University
Abstract
Understanding speech in background noise remains a primary challenge faced by hearing-impaired listeners. Ideal binary masking (IBM) is an effective technique to facilitate understanding of a target signal in noisy backgrounds, and IBM estimation is the goal of an effective speech-from-noise separation algorithm that holds promise for alleviating limitations of hearing impairment. In IBM processing, a speech-and-noise mixture is divided into a grid of time-frequency (T-F) units, which are discarded if their degree of noise corruption (reflected as a signal-to-noise ratio, or SNR) exceeds a certain local criterion (LC). Prior work determined that the relationship between the overall SNR of the original speech-noise mixture and LC (the relative criterion or RC) was important for determining intelligibility. This work also suggests that there is a wide range of RC values over which performance scores reach maximum. The current study investigates whether these scores reflect a performance ceiling rather than a true maximum. Consonant recognition was tested in normal-hearing listeners using seven different RC values. The background noise was speech-shaped noise. An RC performance function was obtained that did not display the ceiling effect limitations of previous work, suggesting that the optimal RC value may be different from previous estimates. These findings have implications for selections of LC in IBM estimations. They also suggest appropriate parameters for testing the effect of varying LC within a single mask according to specific frequency contributions to overall speech intelligibility. Such developments may contribute to reducing the struggles that hearing-impaired listeners face in noise.
Description
1st place winner in the Business/Education/Speech and Hearing Science category in the Denman Undergraduate Research Forum
Keywords
ideal binary mask, hearing science, noise tolerance, speech-from-noise segregation