ISCA Archive SWAP 2000
ISCA Archive SWAP 2000

Context effects on sensitivity and response bias

John Kingston

The massive acoustic redundancy with which minimal contrasts are conveyed and the perceptual integration of these acoustic properties ensure that the listener can identify the intended speech sound from bottom-up information alone most of the time (Kingston, 1991; Kingston & Diehl, 1994, 1995; Kingston & Macmillan, 1995; Kingston, et al., 1997; Macmillan, et al., 1999). The identification of speech sounds is, however, systematically distorted by context (Repp, 1982). These distortions are all contrastive. That is, if the context sound lies at one extreme of an acoustic continuum, the listener gives more responses corresponding to the category at the opposite extreme of the continuum. The criterion shifts both when the extreme value of the acoustic property that varies continuously in the target sound can actually be heard in context sound and when it can only be predicted. When the extreme value of the acoustic property is audible in the context, it creates the contrast regardless of whether it occurs in speech sounds (Mann, 1980, Mann & Repp, 1981) or acoustically analogous nonspeech sounds (Lotto & Kluender, 1998). And when the extreme value can only be predicted, it creates the contrast regardless of whether it's predicted from simultaneous visual information (Fowler, et al., 1999), transitional probabilities (Pitt & McQueen, 1998), or stimulus blocking (Ohala & Feder, 1994; Bradlow & Kingston, 1990).

The audible context effects probably have a different source than the predictable ones. When the extreme value of the acoustic property is audible in the context, it interacts auditorily with its value in the target sound, such that the listener literally hears a contrastive different value in the target. But when the extreme value in the context is predictable, the listener hears the same value in the target sound as when the context isn't present, but shifts the criterion to take into account simultaneous visual, statistical, or inferential information. Audible context effects therefore change the listener's sensitivity as well as response bias to differences in the acoustic property being judged, but predictable context effects change only response bias.

We have recently devised a quantitative model relating sensitivity and bias measures in the fixed classification and trading relations tasks that we can apply to studying these two kinds of context effects (Macmillan, et al., 1999). In the talk I will give at SWAP, I will present the results of an ongoing study comparing the effects of audible vs predictable extreme context values on a well-studied case, the judgment of /t-k/ in the context of /s: /. Two properties that are known to be acoustic and perceptual correlates of the stop place contrast, spectral compactness and transition duration, are varied orthogonally, and listeners both categorize and discriminate the stimuli. Four preceding context conditions are compared, two audible and two predictable. Audible: (1) speech /s/ vs / /, (2) nonspeech, dense, inharmonic tone complexes with the same center frequencies and bandwidths as /s/ and / /. Predictable: The fricative is ambiguous between /s/ and / / but predictable from (3) transitional probabilities between it and the preceding vowel or (4) from the quality of the fricative in the immediately preceding three trials. Conditions (1) and (3) replicate and extend the studies by Mann & Repp and Pitt & McQueen, respectively; condition (2) extends the study of the very similar effect on /d-g/ judgments of /l/ vs /r/ or sine waves mimicking their F3 values by Lotto & Kluender; and condition (4) is entirely novel.


Cite as: Kingston, J. (2000) Context effects on sensitivity and response bias. Proc. Spoken Word Access Processes (SWAP), 151-154

@inproceedings{kingston00_swap,
  author={John Kingston},
  title={{Context effects on sensitivity and response bias}},
  year=2000,
  booktitle={Proc. Spoken Word Access Processes (SWAP)},
  pages={151--154}
}