The Segmentation of Speech

This paper reports a phenomenon supporting the hypothesis that the emergence of structure in the evolution of language was a staged process. To develop a grammatical structure it seems necessary to first have discrete constituents which can be the building blocks of a hierarchical system. By analysing observed speech we show that the development of a linear sequence of grammatical constituents has its own advantage, before a possible next stage when constituents are integrated into a hierarchical structure. A stream of speech sounds has to be segmented to allow for breathing. This segmentation has further developed in a certain way that makes it easier for the hearer to decode than if it were not segmented, or if it were segmented in an arbitrary manner. Well known tools from Information Theory are employed to analyse the ease of decoding speech. Segmentation depends on prosodic discontinuities, such as pauses and intonation marked by tone unit boundaries. These discontinuities usually mark groups of words with some syntactic cohesion, such as phrases and clauses. We show that in a modern corpus of spoken language observed segmentation facilitates the effective transfer of information, while lack of segmentation or arbitrary segmentation imposed on a stream of words makes decoding less efficient. This supports the hypothesis that the necessary constituents of a grammatical structure may have evolved as a consequence of developments favouring more efficient decoding of a linear stream of spoken words. The source material for this investigation is taken from the prosodically marked up Machine Readable Spoken English Corpus (MARSEC).