Perceptual development of phoneme contrasts: how sensitivity changes along acoustic dimensions that contrast phoneme categories.

Listeners discriminate acoustic differences between phoneme categories at a higher level than similarly sized differences within phoneme categories. The question this paper aims to answer is how this pattern in perceptual sensitivity develops along an acoustic dimension that contrasts two non-native speech sounds: through acquired distinctiveness, through acquired similarity, or through a combination of the two. A pretest-training-post-test experiment was designed to study perceptual development directly, i.e., by including (i) a discrimination task to measure perceptual sensitivity, (ii) a transfer test to ensure language learning instead of stimulus learning, and (iii) a control group to exclude task repetition as an explanation of improvement. It is shown that the typical peak in perceptual sensitivity near a phoneme boundary that native listeners show is not found in relatively inexperienced language learners, despite their ability to classify a continuum in a nativelike way after short laboratory training. Experiment II indicates that a discrimination peak may be achieved by language learners, but only after much more language experience than short-term laboratory training can offer. Furthermore, reasons are given why classification improvement in the laboratory should not be taken as evidence for (i) increased discrimination of the newly learned phonemes and (ii) learning of phoneme representations.

Categorical perception depends on the discrimination task.

Speech sounds are said to be perceived categorically. This notion is usually operationalized as the extent to which discrimination of stimuli is predictable from phoneme classification of the same stimuli. In this article, vowel continua were presented to listeners in a four-interval discrimination task (2IFC with flankers, or 4I2AFC) and a classification task. The results showed that there was no indication of categorical perception at all, since observed discrimination was found not to be predictable from the classification data. Variation in design, such as different step sizes or longer interstimulus intervals, did not affect this outcome, but a 2IFC experiment (without flankers, or 2I2AFC) involving the same stimuli elicited the traditional categorical results. These results indicate that the four-interval task made it difficult for listeners to use phonetic information and, hence, that categorical perception may be a function of the type of task used for discrimination.

Searching for an explanation for diphthong perception: dynamic tones and dynamic spectral profiles.

The aim was to find a psychophysical explanation for the perception, by naive listeners, of diphthongs as single vowels, even though they are essentially formant movements. Subjects were asked to match sinusoidal tone and resonance glides around 1,000 Hz with two connected steady-state tones or resonances whose frequencies could be controlled independently. The expectation was that short glides (below 120 ms) would give rise to single perceptual events without any movement in a particular direction, so that the two matching steady-state patterns would not show any frequency direction either; long resonance glides (above 120 ms), on the other hand, were expected to be perceived as rising or falling and matched accordingly. The results showed an effect of duration, although it interacted with glide width. At durations shorter than about 120 ms, subjects placed the two steady profiles with which they had to match the dynamic profile closer together than with durations over 120 ms; however, this only occurred if a glide covered more than 500 Hz, and is therefore irrelevant to diphthong perception.

Categorical perception as a function of stimulus quality.

A number of experiments were carried out in order to test the hypothesis that categorical perception of speech stimuli is a function of synthesis quality - specifically, that the greater complexity of more natural speech stimuli makes it difficult for listeners to focus on particular stimulus parameters as psychoacoustic cues. The results show that there is an increase in categorical perception as synthesis quality improves from a simple synthesis by rule, via LPC synthesis, to a much more complex type called sinewave generation.

Response distributions in intensity resolution and speech discrimination.

In this paper the assumption of an equal, Gaussian distribution of the response to each stimulus in an experiment, an assumption which has to be met if d' is to be estimated by calculating the difference between z(H) and z(FA), is tested for two different sets of stimuli: 1000-Hz tones differing in level only, and a continuum of stop consonants, obtained by full spectral interpolation between /p/, /t/, and /k/. Response distributions were measured directly by means of a form of non-numerical magnitude estimation, in which subjects had to indicate the position of each stimulus on a quasi-continuous rating scale. It could be shown that, in general, all distributions were sufficiently unimodal, but that their variances differed. The consequences for the calculation of d' are unlikely to be serious.

Modeling phoneme perception. I: Categorical perception.

On the basis of a number of vowel and stop-consonant discrimination experiments (AX and 2IFC fixed and roving) with natural stimuli, it is concluded that stop-consonant perception is highly categorical: there were few significant differences between the discrimination results and the phoneme identification results. Moreover, the discrimination and identification response maxima differed significantly from the other data points. Vowel perception was much less categorical: the maxima in the functions were much less significant, and there were significant differences between the various paradigms. Consonant discrimination was much less (if at all) subject to range effects than vowel discrimination. All these results point to different memory types for stop consonants and vowels, and, consequently, to a combination of two different theories of speech sound discrimination: dual-process theory (DPT) for consonants, and trace-context theory (TCT) for vowels.

Modeling phoneme perception. II: A model of stop consonant discrimination.

Combining elements from two existing theories of speech sound discrimination, dual process theory (DPT) and trace context theory (TCT), a new theory, called phoneme perception theory, is proposed, consisting of a long-term phoneme memory, a context-coding memory, and a trace memory, each with its own time constants. This theory is tested by means of stop-consonant discrimination data in which interstimulus interval (ISI; values of 100, 300, and 2000 ms) is an important variable. It is shown that discrimination in which labeling plays an important part (2IFC and AX between category) benefits from increased ISI, whereas discrimination in which only sensory traces are compared (AX within category), decreases with increasing ISI. The theory is also tested on speech discrimination data from the literature in which ISI is a variable [Pisoni, J. Acoust. Soc. Am. 36, 277-282 (1964); Cowan and Morse, J. Acoust. Soc. Am. 79, 500-507 (1986)]. It is concluded that the number of parameters in trace context theory is not sufficient to account for most speech-sound discrimination data and that a few additional assumptions are needed, such as a form of sublabeling, in which subjects encode the quality of a stimulus as a member of a category, and which requires processing time.

Identification and discrimination of sweep formants.

Earlier identification experiments with sweep tones are repeated with rising and falling single formant (band) sweeps, with durations ranging from 15 to 40 msec and sweep rates from 0 to 40 oct/sec. Steady-state portions of 100-msec duration are then added to the sweeps. The general conclusions are that the tendency to perceive level and slightly rising tones as falling, which was such a prominent feature of the earlier results, disappears as the stimuli become more complex, and that sweep discrimination seems to be a function of the difference between the initial and the final frequency of a sweep.

Identification of deleted plosives: The effect of adding noise or applying a time window (a reply of Ohde and sharf).

A Letter in this Journal [Ohde, R. N. and Sharf, D. J., J. Acoust. Soc. Am. 69, 297-300 (1981)] deals with the identification of voiced and voiceless stops from vocalic transition + vowel stimuli edited from CV and VC syllables. The segmentation points were smoothed by applying a time window. With respect to an earlier study, in which straightforward segmentation without time window was applied, final plosives were still identified better than initial plosives. The present Letter shows that there is, however, one major difference between the with-window and the no-window condition, namely a substantial improvement in identification for initial voiced plosives. This improvement is similar to that found by the present authors in identification experiments with a noise burst adjacent to the segmentation point versus no-noise listening conditions. Instead of a controversy reported by Ohde and Sharf, this Letter suggests a unifying interpretation.

Identification of deleted consonants.

VC transitions have often been found to be better cues for the identification of deleted consonants than CV transitions. The lower score for CV transitions could very well be due to an interfering "click sensation" caused by the abrupt begining of these transitions. In the present investigation the abrubt onset was eliminated by replacing the deleted portions of the consonants with noise bursts; this also caused the incomplete syllables to sound more natural. The results show that the identification of deleted initial voiceless plosives is greatly improved by the addition of noise. The original difference between initial and final plosive transitions disappears almost completely.