Automatic context-sensitive measurement of the acoustic correlates of distinctive features at landmarks

This paper models speech recognition as the estimation of distinctive feature values at articulatory landmarks 8]. Toward this end, we propose modeling each distinctive feature as a table containing phonetic contexts, a list of signal measurements (acoustic correlates) which provide information about the feature in each context, and, for each context, a statistical model for evaluating the feature given the measurements. The model of a distinctive feature may include several sets of acoustic correlates, each indexed by a diierent set of context features. Context features are typically lower-level features of the same segment, e.g. manner features ((con-tinuant, sonorant]) provide context for the identiication of articulator-bound features ((lips, blade]). The acoustic correlates of a feature can be any static or dynamic spectral measurements deened relative to the time of the landmark. The statistical model is a simple N-dimensional Gaussian hypothesis test. A measurement program has been developed to test the usefulness of user-deened acoustic correlates in user-deened phonetic contexts. Measures of voice onset time and formant locus classiication are presented as examples. 1 Recognition of planning units When a word changes tense, case, or number, the phonemes in the word often change in predictable ways. Distinctive feature notation was developed as a compact notation for these empirical rules of phonological alternation. As such, distinctive features are a model of a data representation used by the human brain to plan speech production 2]. In order to be useful for modeling speech production, a set of distinctive features must be as compact as possible, while providing unique representations for all distinguish-able allophones. In English, for example, there are about forty phonemes, and one to two hundred allophones, all of which may be uniquely represented in terms of about twenty binary distinctive features. In this paper, we will use a binary articulatory feature set capable of representing some syllabic and prosodic information, in addition to traditional allophone distinctions ((gure 1). Since there are fewer distinctive features than phonemes, and since they provide a more parsimonious representation of coarticulation, automatic recognition of distinctive features should be more eecient than automatic recognition of