Speech Processing In Adverse Conditions

Speech recognition in ‘adverse conditions’ has been a familiar area of research in computer science, engineering, and hearing sciences for several decades. In contrast, most psycholinguistic theories of speech recognition are built upon evidence gathered from tasks performed by healthy listeners on carefully recorded speech, in a quiet environment, and under conditions of undivided attention. Building upon the momentum initiated by the Psycholinguistic Approaches to Speech Recognition in Adverse Conditions workshop held in Bristol, UK, in 2010, the aim of this volume is to promote a multi-disciplinary, yet unified approach to the perceptual, cognitive, and neuro-physiological mechanisms underpinning the recognition of degraded speech, variable speech, speech experienced under cognitive load, and speech experienced by theoretically relevant populations. This collection opens with a review of the literature and a formal classification of adverse conditions. The research articles then highlight those adverse conditions with the greatest potential for constraining theory, showing that some speech phenomena often believed to be immutable can be affected by noise, surface variations, or attentional set in ways that will force researchers to rethink their theory. This volume is essential for those interested in speech recognition outside laboratory constraints.

Although speech is the primary behavioral medium by which humans communicate, its auditory basis is poorly understood, having profound implications on efforts to ameliorate the behavioral consequences of hearing impairment and on the development of robust algorithms for computer speech recognition. In this volume, the authors provide an up-to-date synthesis of recent research in the area of speech processing in the auditory system, bringing together a diverse range of scientists to present the subject from an interdisciplinary perspective. Of particular concern is the ability to understand speech in uncertain, potentially adverse acoustic environments, currently the bane of both hearing aid and speech recognition technology. There is increasing evidence that the perceptual stability characteristic of speech understanding is due, at least in part, to elegant transformations of the acoustic signal performed by auditory mechanisms. As a comprehensive review of speech's auditory basis, this book will interest physiologists, anatomists, psychologists, phoneticians, computer scientists, biomedical and electrical engineers, and clinicians.

This thesis investigated subjective, behavioural and neurophysiological (EEG) measures of speech processing in various adverse conditions and with different listener groups. In particular, this thesis focused on different neural processing stages and their relationship with auditory attention, effort, and measures of speech intelligibility. Study 1 set the groundwork by establishing a toolbox of various neural measures to investigate online speech processing, from the frequency following response (FFR) and cortical measures of speech processing, to the N400, a measure of lexico-semantic processing. Results showed that peripheral processing is heavily influenced by stimulus characteristics such as degradation, whereas central processing units are more closely linked to higher-order phenomena such as speech intelligibility. In Study 2, a similar experimental paradigm was used to investigate differences in neural processing between a hearing-impaired and a normal-hearing group. Subjects were presented with short stories in different levels of multi-talker babble noise, and with different settings on their hearing aids. Findings indicate that, particularly at lower noise levels, the hearing-impaired group showed much higher cortical entrainment than the normal- hearing group, despite similar levels of speech recognition. Intersubject correlation, another global neural measure of auditory attention, however, was similarly affected by noise levels in both the hearing-impaired and the normal-hearing group. This finding indicates extra processing in the hearing-impaired group only on the level of the auditory cortex. Study 3, in contrast to Studies 1 and 2 (which both investigated the effects of bottom-up factors on neural processing), examined the links between entrainment and top-down factors, specifically motivation; as well as reasons for the 5 higher entrainment found in hearing-impaired subjects in Study 2. Results indicated that, while behaviourally there was no difference between incentive and non-incentive conditions, neurophysiological measures of attention such as intersubject correlation were affected by the presence of an incentive to perform better. Moreover, using a specific degradation type resulted in subjects' increased cortical entrainment under degraded conditions. These findings support the hypothesis that top-down factors such as motivation influence neurophysiological measures; and that higher entrainment to degraded speech might be triggered specifically by the reduced availability of spectral detail contained in speech.

The domain of speech processing has come to the point where researchers and engineers are concerned with how speech technology can be applied to new products, and how this technology will transform our future. One important problem is to improve robustness of speech processing under adverse conditions, which is the subject of this book. Robust speech processing is a relatively new area which became a concern as technology started moving from laboratory to field applications. A method or an algorithm is robust if it can deal with a broad range of applications and adapt to unknown conditions. Robustness in Automatic Speech Recognition addresses all of the fundamental problems and issues in the area. The book is divided into three parts. The first provides the background necessary for understanding the rest of the material. It also emphasizes the problems of speech production and perception in noise along with popular techniques used in speech analysis and automatic speech recognition. Part Two discusses the problems relevant to robustness in automatic speech recognition and speech-based applications. It emphasizes intra- and inter-speaker variability as well as automatic speech recognition of Lombard, noisy and channel distorted speech. Finally, the third part covers recent advances in the field of robust automatic speech recognition. Audience: An invaluable reference. May be used as a text for advanced courses on the subject.

Proceedings of the NATO Advanced Study Institute on Computational Models of Speech Pattern Processing, held in St. Helier, Jersey, UK, July 7-18, 1997