Representational Maps from the Speech Signal to Phonological Categories: a Case Study with Lexical Tones

As the initial step in studying the acquisition of phonological categories from the speech signal, we describe representational issues for the target of learning, a probabilistic distribution of phonological categories over a phonetic parameter space. Our model system of study is cross-linguistic lexical tonal phonemes in tonal languages. We focus on two representational issues: temporal resolution of the extracted phonetic parameters and static and dynamic parameterizations of the speech signal. In a human perception study and exploratory computational modeling, we find that coarse sampling of absolute f0 and f0 velocity is sufficient for near-partitions of the phonetic parameter space for single-speaker tonal spaces in a range of tone languages.