Acoustic cues in the directionality of stop consonant confusions

Stop place identification studies of consonant vowel sequences (CVs) in English (and many other languages) report similar patterns: Listener errors vary by (1) stop place, (2) the following vowel, and (3) confusion direction. Asymmetries in the direction of stop place confusions (e.g., [ki] is often confused for [ti], but [ti] is rarely confused for [ki]) have been noted by many researchers, but not yet adequately explained. This perceptual study evaluates previous accounts for asymmetry in stop place confusions and tests a new hypothesis: that listeners may disfavor stop places with greater token-to-token variation along a primary cue over those with less variation.;The current study differs in methodology from classical perceptual studies, in which generally one or two acoustic properties in synthetic or natural speech known to cue stop place are carefully manipulated. Instead, a combination of digital signal processing and machine learning techniques are used to examine a larger set of stop place cues to determine how they are ranked as well as how they vary by context and stop place. Additionally, the current study uses the natural token-to-token variation of speech to avoid any artifacts that artificial or manipulated speech may cause in listener perceptions.;Unaltered CV tokens (C = {p, t, k} and V = {i, a, u}) were collected from careful speech of native English speakers. Several potential cues to stop place were extracted for each token, including the rate and transition of the second formant during vowel onset, the gross spectral shape and relative amplitude of the stop burst, and voice onset time (VOT). CV tokens representing all nine contexts were chosen to serve as stimuli in a perception experiment. The unaltered stimuli were presented to experimental subjects for stop place identification in a cross-modal task designed to induce errors.;CVs that are more frequent in a listener's lexicon were more frequently given as incorrect responses in the stop place identification task. Additionally, perceptual asymmetry was caused by differences in token-to-token variation. In certain CV contexts, stimuli with ambiguous values were significantly more likely to be confused than those with canonical values. Phonotactic frequency in the lexicon, degradation of non-robust cues, and differences in production variation all affect perceptual error rates differentially by CV context. The results show that humans employ whatever cues (phonetic or non-phonetic) are available for the categorization of speech sounds, especially when primary acoustic cues are ambiguous.