Significant features in the perception of (Hindi) consonants

Analysis of Spanish consonant recognition in 8-talker babble

This paper presents the results of a closed-set recognition task for 80 Spanish consonant-vowel sounds (16 C × 5 V, spoken by 2 talkers) in 8-talker babble (-6, -2, +2 dB). A ranking of resistance to noise was obtained using the signal detection d' measure, and confusion patterns were analyzed using a graphical method (confusion graphs). The resulting ranking indicated the existence of three resistance groups: (1) high resistance: /ʧ, s, ʝ/; (2) mid resistance: /r, l, m, n/; and (3) low resistance: /t, θ, x, ɡ, b, d, k, f, p/. Confusions involved mostly place of articulation and voicing errors, and occurred especially among consonants in the same resistance group. Three perceptual confusion groups were identified: the three low-energy fricatives (i.e., /f, θ, x/), the six stops (i.e., /p, t, k, b, d, ɡ/), and three consonants with clear formant structure (i.e., /m, n, l/). The factors underlying consonant resistance and confusion patterns are discussed. The results are compared with data from other languages.

The role of perception in the sound change of velar palatalization

Voiceless velar stops may become palatoalveolar affricates before front vowels. This sound change is not only one of the most common types of palatalization, but is a very common sound change in the world's languages. Nevertheless, we do not have an adequate understanding of how this sound change takes place. Three experiments reported here test the hypothesis that velar palatalization is the result of listeners' on-line perceptual reanalysis of fast rate speech. It is shown that velars before front vowels are both acoustically and perceptually similar to palatoalveolars. This supports the proposal that velar palatalization is perceptually conditioned.

An investigation of stop place of articulation as a function of syllable position: a locus equation perspective

Locus equations were employed to phonetically describe stop place categories as a function of syllable-initial, -medial, and -final position. Ten speakers, five male and five female, produced a total of 2700 CVC and 4500 VCV utterances that were acoustically analyzed to obtain F2 onset, F2 vowel, and F2 offset frequencies for locus equation regression analyses. In general, degree of coarticulation, as indexed by locus equation slope, was reduced for post-vocalic (VC) stops relative to pre-vocalic stops (pooled data from initial and medial positions), but significant differences were observed as a function of stop consonant. All stops showed significantly reduced R2 values and increased standard errors of estimate for VC relative to CV productions. Separability of stop place categories in a higher-order slope X y-intercept acoustic space also diminished for VC vs CV stop productions. The degradation of classic locus equation form (high correlation and linearity) for VC relative to CV productions was attributed to greater articulatory precision in the production of pre-vocalic compared to post-vocalic stops. This greater articulatory precision was interpreted as reflecting a greater need to normalize vowel context-induced variability of the F2 transition for syllable onset relative to final stops. The decline in acoustic lawfulness of syllable-final stops is discussed in terms of coarticulatory interactions and expected perceptual correlates.

Feature-testing models and their application to perception and memory for speech

Four models of the organization of feature-testing units in the perception of phonemes are developed and tested against data drawn from a number of published studies. The independence model, which asserts that features are tested independently of each other, is conclusively rejected. The response-conditional and stimulus-conditional models require individual feature-testing units to be dependent, respectively, on the responses of other feature-testing units and on the other features present in the stimulus. Both these models handle perceptual confusions very well among stop and nasal consonants, but are less satisfactory for fricatives. Finally, the correct-conditional model requires individual feature-testing units to be dependent on the correctness of the decisions made by other feature-testing units; this model is very poor for perceptual confusions but quite good for mnemonic confusions. In addition, some of the distinctive features currently favoured by linguists are shown to provide unsatisfactory descriptions.

The implications of this study are that perceptual and mnemonic processes are different in certain basic respects, and that many current scaling models are ignoring important types of structure in the data they analyse.