Participants in this study were trained to categorize a 7-token continuum of unfamiliar nonphonemic sounds with speech-like (i.e., human voice) acoustic properties into two discrete categories (A and B).
Under the logistic regression framework, the probability of a certain response (category 1) can be modeled as P category 1) = 1/(1 + e−(α + β x )), where X is the predictor variable (here, the position of the token in the continuum).
According to Gordon [15], vowel devoicing operates on a continuum with token to token variation; on one end of the spectrum is a devoiced vowel and on the other end is vowel deletion, with various degrees of voicing falling in between.
In 4 subsequent training sessions occurring over the course of approximately 2 weeks, participants practiced categorizing all 7 tokens in each continuum into the two categories represented by the end points, through gradual introduction of token pairs closer to the category boundary.
Participants were then scanned while performing the categorization task with all the tokens in each continuum.
These were stimulus tokens 1, 7, and 14 from the /ba, da, ga/ continuum, and tokens 1, 7, and 13 from the /bu, du, gu/ continuum, respectively.
However, firstly, such natural productions involved an in-depth knowledge of the relevant articulatory changes that are required for creating the uniform CoG steps; secondly, selecting the right tokens for the CoG continuum involved phonetically trained ears.
P tokens were composed of a continuum from /ba/ to /da/.
The sibilants correspond to tokens 3 12 of the original sibilant continuum; b more detailed frequency analysis of the two sibilants 1 and 10 (formerly 3 and 12) at the endpoints of the continuum.
In the pretraining session before scanning, participants were first introduced only to the end tokens (1 and 7) of each continuum and briefly practiced to categorize them as /ba/ or /da/ (P continuum) and A or B (NP continuum), using 1 of 2 keys.
P and NP continua were matched on token duration, amplitude, spectrotemporal complexity, and overall formant structure but differed in that the stop-like onsets of the NP items were unfamiliar and not analogous to any English phoneme.
