Speech Communication from an Information Theoretical Perspective

Throughout the last century, models of human speech communication have been proposed by linguists, psychologists, and engineers. Advancements have been made, but a theory of human speech communication that is both comprehensive and quantitative is yet to emerge. This thesis hypothesises that a branch of mathematics known as information theory holds the answer to a more complete theory. Information theory has made fundamental contributions to wireless communications, computer science, statistical inference, cryptography, thermodynamics, and biology. There is no reason that information theory cannot be applied to human speech communication, but thus far, a relatively small effort has been made to do so.  The goal of this research was to develop a quantitative model of speech communication that is consistent with our knowledge of linguistics and that is accurate enough to predict the intelligibility of speech signals. Specifically, this thesis focuses on the following research questions: 1) how does the acoustic information rate of speech compare to the lexical information rate of speech? 2) How can information theory be used to predict the intelligibility of speech-based communication systems? 3) How well do competing models of speech communication predict intelligibility?  To answer the first research question, novel approaches for estimating the information rate of speech communication are proposed. Unlike existing approaches, the methods proposed in this thesis rely on having a chorus of speech signals where each signal in the chorus contains the same linguistic message, but is spoken by a different talker. The advantage of this approach is that variability inherent in the production of speech can be accounted for. The approach gives an estimate of about 180 b/s. This is three times larger than estimates based on lexical models, but it is an order of magnitude smaller than previous estimates that rely on acoustic signals.  To answer the second research question, a novel instrumental intelligibility metric called speech intelligibility in bits (SIIB) and a variant called SIIBGauss are proposed. SIIB is an estimate of the amount of information shared between a talker and a listener in bits per second. Unlike existing intelligibility metrics that are based on information theory, SIIB accounts for talker variability and statistical dependencies between time-frequency units.   Finally, to answer the third research question, a comprehensive evaluation of intrusive intelligibility metrics is provided. The results show that SIIB and SIIBGauss have state-of-the-art performance, that intelligibility metrics tend to perform poorly on data sets that were not used during their development, and show the advantage of reducing statistical dependencies between input features.