Andrew Radford - Linguistics An Introduction [Second Edition]

An alternative perspective is, however, presented by our discussion in section 5, where we saw that as far as the structure of the language is concerned, this infinity of speech sounds is reduced to a finite inventory of functioning units, the phonemes of the language.

Let’s approach this topic by changing our question. Rather than being con-

cerned with what speech sounds are like, let’s ask what our perception of speech sounds is like? Obviously, it could be the case that we perceive all the infinite gradations which the continuous nature of such notions of place of articulation, front/back and high/low make available, or it might be that our perceptual systems are ‘tuned’ to the phonological structure of our native language, so that we simply do not hear differences in speech sounds which are not linguistically significant.

The answer to our revised question is surprisingly complex, and it is likely that a complete understanding of this matter lies some way in the future. Part of the answer, however, seems to depend on what sort of speech sound we are

considering.

In order to investigate systematically the issue which concerns us, it is important to be able to control the characteristics of the speech sounds we test. Native speakers cannot vary their speech sounds with the required degree of control, but it is possible to produce speech sounds synthetically using a speech synthesiser.

For example, reasonably accurate tokens of syllables such as /ba/ or /pe/ can be produced in this way, and it is then possible to introduce slight, carefully controlled changes into the acoustic form of the synthesised syllables and words, changes which correspond to a gradual shift in place of articulation of a consonant or the height or frontness of a vowel, etc.

One set of experiments we can perform is on vowel sounds. We can synthesise

tokens of, say, the words pit and pet. Then, starting with our token of pit, we change its acoustic characteristics in a number of discrete steps until we get to our token of pet. The outcome of this process is referred to as an [i – ɛ]-series, i.e. a set of synthesised stimuli with something which is unambiguously pit at one end, something which is unambiguously pet at the other and a number of acoustically intermediate forms. Such a series can then be used in a variety of experiments with native speakers.

One commonly performed experiment is an identification experiment. In this,

members of the series are simply presented to native speakers in random order, and they have to say whether they hear pit or pet – note that we do not allow them to say that a stimulus is neither pit nor pet, i.e. we employ what is called a ‘forced choice paradigm’. A typical (idealised) result from such an experiment appears in

figure 34.

Here, along the y-axis, we have the number of times subjects report that they have heard pit as opposed to pet, and along the x-axis the items in the series of synthesised stimuli, with 1 corresponding to the original pit, 10 to the original pet

Processing sounds

111

Figure 34 Results of an identification experiment for an [ɪ – ɛ]-series

Figure 35 Results of a discrimination experiment for an [ɪ – ɛ]-series

and 2–9 labelling the intermediate stimuli. What results such as this seem to show is that the perception of vowels is continuous, with each vowel appearing to shade gradually into the next. For items such as 4, 5 and 6, intermediate between pit and pet, subjects appear to have recourse to guessing.

A rather different experimental procedure which leads to the same conclusion is a discrimination experiment. Such an experiment typically presents native

speakers with pairs of adjacent stimuli from a synthesised series followed by a third stimulus which is identical to one of the first two. The subjects’ task is to say whether the third stimulus is identical to the first or the second. Obviously, we would expect such a task to be difficult for subjects if their perception is continuous, and this turns out to be the case for a vowel series such as that considered above. Results of a typical experiment are presented in figure 35 (again these are idealised – empirical enquiry never yields lines as straight as this – but this does not affect the point under discussion).

Here, on the x-axis, we have pairs of synthesised stimuli which are presented for discrimination, and what the straight line indicates is that subjects did only slightly better across the whole series than they would if they were guessing, i.e. discrimination of adjacent pairs was uniformly poor in this case.

What we have described so far is perhaps not very surprising, but when we turn to the perception of consonants, a very different picture emerges. A contrast which has been extensively studied is the voiced–voiceless contrast in [b/p], [d/t] and

112

sounds

so on. As we know from section 2, voicing occurs when the state of the larynx permits the vocal cords to vibrate. In our earlier discussion, we talked as if voicing occurs during the production of voiced consonants, but for plosives this is not quite correct. In fact, if the syllable [ba] is produced in English, the vocal cords do not begin to vibrate until a short time after the release of the bilabial closure. By contrast, if [pa] is produced, there is a relatively long time between the release of the closure at the lips and the onset of vocal cord vibration for the vowel, and if the consonant is heavily aspirated, this time becomes even longer. Thus, the acoustic correlate of the distinction between voiced, voiceless and aspirated voiceless plosives lies in the time interval between the release of the closure and the beginning of the voicing associated with the following vowel sound. This interval is called Voice Onset Time or VOT. Now, of course, time is a continuous variable, and using synthetic stimuli, it is possible to create a set of syllables, comprising a [b–p]-series in which VOT is systematically varied. Obviously, with a short VOT, we expect subjects to perceive [b], whereas with a long VOT, we predict that they will perceive [p]. The interesting question is what happens with intermediate values?

In figure 36 we see the results of an identification experiment on the perception of [b] and [p], with VOT varying along the x-axis.

What is significant here is what happens when the VOT value is about 25 ms.

Subjects shift suddenly from reporting [b] to reporting [p]. However, any VOT

value less than about 20 ms. is heard as [b], while any VOT greater than about 30 ms. is reported as [p].

Of course, on the basis of this identification experiment, we cannot conclude anything about the subjects’ abilities to perceive distinctions within categories, but the discrimination experiment enables us to investigate this. What we find here is that if test stimuli fall on opposite sides of the boundary indicated in figure 36, subjects are very accurate in their identifications. If, however, the stimuli fall on the same side of the boundary, then subjects’ responses indicate that they are guessing, i.e. they cannot perceive the difference between a stimulus with a VOT

of, say, 40 ms. and another with a VOT of, say, 60 ms. Typical results from such Figure 36 Results of an identification experiment for a [b – p]-series

Processing sounds

113

Figure 37 Results of a discrimination experiment for a [b – p]-series

an experiment appear in figure 37, where on the x-axis we have the mean VOTs for the stimuli being presented for discrimination (e.g. 25 ms. gives the result for discrimination of stimuli with 20 ms. and 30 ms. VOT).

What figure 37 indicates is that hearers can discriminate the phonetic categories, voiced v. voiceless, very well but they cannot hear differences within these categories. This type of perception is known as categorical perception because the hearer perceives in terms of categories (voiced or voiceless) rather than in terms of minute gradations of sound. It is of considerable interest that