Key Issues and Future Directions: Models of Human Language and Speech Processing
Willem H. Zuidema, Hartmut Fitz · The MIT Press eBooks · 2019
FITZ of vibrating air.The inner ear converts the stream of input into spike trains in the auditory nerve, and the receiver's brain somehow discovers (or imposes) segments in the input, recognizes segments as members of a category (phonemes, syllables, words), recognizes relations between segments in the input, assigns meaning to them, and decides whether the input is wellformed, incomplete, or ungrammatical.In modeling the neural and cognitive pro cesses involved in interpreting a spoken utterance (and similarly, in production, acquisition, and evolution of written or signed languages), modelers have to make a series of choices and simplify the unorderly, complex real ity.Do we focus on the physical real ity of the speech signal, on the neural real ity of pro cessing in the brain, or on the psychological real ity of understanding a received message?The dif fer ent research traditions reviewed in the previous chapters-including symbolic and neural network modeling paradigms that are often presented as incompatible-start from dif fer ent answers to these questions.For instance, the models of syntax and sentence processing reviewed in Demberg and Keller (chapter 22 of this volume), as well as many of the models of language generation reviewed in Krahmer (chapter 25) take abstract syntactic categories and hierarchical structure as a starting point, while greatly simplifying the nature of the signals.The vari ous neural network models discussed in Frank, Monaghan, and Tsoukala (chapter 21) and Zuidema and Le (chapter 23), on the other hand, aim to account for how the empirical observations of linguistic be hav ior, with what at least to some extent looks like discrete categories and hierarchy, might emerge from the interaction between nodes with continuous activation values in a network.The work discussed in Wehbe, Fyshe, and Mitchell (chapter 24), then, addresses explic itly the relation of these kind of models with detectable activity in the human brain.Fi nally, the models of speech production and the vocal