The size and generality of units in speech perception

Abstract

Listen

Translate article

In English, the perception of syllables can be largely predicted from the perception of phonemes. Experiments reviewed by Allen [IEEE Trans. Speech and Aud. Proc. 2, 567–577] have consistently shown that the correct identification of nonsense CVC syllables in noise is extremely well predicted from the marginal identification rates of their constituent phonemes. Simulations suggest that this result can be easily achieved when syllable patterns can be ’factored’ into phoneme parts. Representations allowing even slightly idiosyncratic relationships between stimuli and syllables apparently cannot produce such results. Evidence for phoneme factorability is also provided by the results of certain parametric speech perception experiments. In such experiments, listeners hear synthetic syllables that span two or more categories of two or more segments (e.g., ‘‘bad,’’ ‘‘bet,’’ ‘‘bat,’’ ‘‘bet’’). Analysis of several such experiments shows that syllable identification can be accurately modeled using independent phoneme-sized units that are sensitive to stimulus properties, plus certain highly constrained influences of higher level symbolic elements (e.g., words). These higher level influences are limited to bias effects that do not interact with stimulus properties. These results suggest that small, highly general elements play the key role in the transduction of signals to symbol in speech. [Work supported by SSHRC.]