Intelligibility Estimation in Law Enforcement Speech Processing

摘要

Speech recordings obtained in the context of law enforcement are often degraded in terms of quality and intelligibility. Several techniques for assessing the impact of speech enhancement algorithms on quality are available, both intrusive and nonintrusive, but the assessment of intelligibility is usually reliant on expensive and time consuming subjective listening scores. To address this issue, we describe some recent scoring experiments and an adaptive Bayesian procedure which efficiently estimates properties of the psychometric function from a small number of listening tests. A data-driven nonintrusive objective intelligibility estimation method is also described and tested on car and babble noise. It is shown to give intelligibility estimates that are well correlated with subjective scores. We aim to improve our understanding of the quality/intelligibility tradeoff and to study speech processing tools in the critical context of law enforcement.