Cortical Processing Approaches to Mitigation of Reverberation in Speech

ABSTRACT: Vocal communication in the real world often takes place in complex and noisy acoustic environments. For applications such as mobile communications or use of automated speech/speaker recognition algorithms on recorded data, the presence of noise and reverberation significantly degrades the quality and intelligibility of speech and severely hampers the performance of these systems. Approaches to mitigate the effects of reverberation thus far have either been too dependent upon exact knowledge of the reverberant environment or have resulted in a severely degraded speech signal. The team of In-Depth Engineering and the University of Maryland proposes to explore the use of a novel biomimetic approach to the decomposition of sound into spectrotemporal modulations, called Cortical Processing, to the problem of mitigating the effects of reverberation on speech communications. Cortical Processing will be used to decompose both pure speech signals and speech signals corrupted by reverberation, identify modulation components unique to each, and then create filters to suppress reverberation and enhance pure speech. Recent research findings regarding dynamic changes in neural response properties in the presence of noise give rise to another aspect of reverberation suppression complimentary to Cortical Processing that will be examined as part of the Phase I effort.; BENEFIT: The benefits of the proposed reverberation mitigation approach over previous methods principally arise from the fact that Cortical Processing (1) does not depend on knowledge of the environment in which the sound is heard or recorded, and (2) because Cortical Processing provides such a large feature space via spectrotemporal sound decomposition, the process of removing noise and reverberation effects and reconstructed the clean speech signal does not severely degrade the intelligibility of the speech. The proposed reverberation mitigation tool would be valuable in multiple domains in addition to the intended transition path with the Air Force for enhancing the performance of automated speech analysis and speaker recognition algorithms. During the Phase I effort, the team will consider several commercialization paths, including application to speech recognition and speech translation technologies for aircraft communications, automated audio analysis and audio classification algorithms, especially for speech analysis and speaker recognition, for the intelligence community, removal of reverberation in hands-free mobile communications for the Army, Navy, and Marines, for applications of automated indoor acoustic monitoring systems for facility security in which reverberation will affect the classifier performance, and the commercial marketplace application to wireless hands-free automobile communications where reverberation is a pervasive problem. In particular, In-Depth will partner with a commercial wireless mobile phone company during Phase I in order to build a potential transition plan.