A mapping between trainable generalized properties and the acoustic correlates of distinctive features

Abstract

It has been suggested that the acoustic correlates of distinctive features may be most reliably measured relative to important acoustic events, or landmarks [K. N. Stevens et al., Proc. Int. Conf. Spoken Language Process. 1, 499–502 (1992)]. Theoretical and empirical literature in the field of phonetics describes many of the acoustic correlates of distinctive features, but these acoustic correlates are often described in terms of formant frequencies and amplitudes, aspiration noise, and other attributes which are difficult for a speech recognizer to measure reliably. This paper reports on a preliminary attempt to compile a table of acoustic correlates of distinctive features, and to map these correlates to reliable acoustic measurements in the vicinity of an acoustic landmark. The acoustic measurements are simple energy and spectral change measurements, defined in terms of frequency and timing parameters which may be optimized for maximum feature discrimination using the technique described by Phillips and Zue [Proc. Int. Conf. Spoken Language Process. 1, 795–798 (1992)]. The table is indexed by distinctive feature, and if an acoustic correlate is likely to be salient only in a given phonological context, the context is given as a list of neighboring and simultaneous distinctive features. [Work supported by NSF.]