Signal densities and criterion variance in speech and nonspeech perception

Abstract

The actual shape of signal densities has become an important issue when studying speech perception within the framework of Signal Detection Theory (SDT). Using an SDT model that allowed unequal criterion variances, López-Bascuas [Proc. Eurosp. 3, 2281-2283 (1995)] found that speech signals did not accommodate to the standard Gaussian assumption. However, Schouten and van Hessen [J. Acoust. Soc. Am. 104, 2980-2990 (1998)] measured response distributions directly and, assuming an interval scale, concluded that the Gaussian assumption held for both continua. Nevertheless, Pastore and Macmillan [J. Acoust. Soc. Am. 111, 2432 (2002)] applied ROC analysis to Schouten and van Hessen's data and their curves supported the Gaussian assumption for the nonspeech signals only. Later, López-Bascuas et al. [J. Acoust. Soc. Am. 115, 2465 (2004)] showed that non-linear z-transformed ROCs are not sufficient evidence for postulating non-gaussian signal densities. In this paper we try to figure out whether unequal criterion variances could underlie non-linear z-transformed ROCs by fitting a restricted Thurstonian SDT-like model to a nonspeech continuum composed by white noise and a square wave. The results indicate that unequal criterion variances is the possible cause of the deviant ROCs for nonspeech.