SPEECH PERCEPTION: PHONEME VARIATION

It has proved difficult for listening researchers to account for our ability to identify the sounds of speech. This is because realisations of a given phoneme vary enormously.

Non-linearity. The spoken signal does not consist of a string of phonemes in the way that written language consists of a string of letters. In any token of the word [k æ t], there is no precise boundary for the sound /k/: co-articulation blends it into the succeeding / æ /, and the / æ / blends into the /t/.

Non-invariance. Co-articulation also means that there is no standard form for any given phoneme. The articulation of a vowel can range over a wide vowel space. Similarly, the place of articulation of a consonant such as /k/ varies greatly anticipating the vowel that is to follow it (compare /k/ in KILL with /k/ in CUT). Researchers have studied the cues that are physically present in the speech stream, but have failed to find any combination of features that uniquely serve to characterise a given consonant. They conclude that listeners recognise a consonant partly through the co-occurrence of several phonetic features (not always the same ones), but in particular through the shape of the formant transition which links a consonant to the following vowel.

Speaker variation. There is enormous between-speaker variation. Every speaker has a distinctive voice: our articulators (mouths, jaws, tongues, teeth) vary greatly in size, shape and position. Many speakers have regional accents. There are important differences in pitch between the voices of men and the voices of women and children. We speak at different rates with different degrees of emphasis and with different idiolectal features. Researchers are interested in how we manage to normalise (adjust) to the voice and speech rate and accent of a wide range of individuals– and what kind of representation we match our perceptions against. We continually make fine judgements based upon rate of articulation: our judgement of how rapidly the speaker is articulating may even influence which phoneme we hear.

Within-speaker variation. There is variation within the speech of a single speaker. For example, speech rate may vary enormously according to mood, auditor and communicative purpose, as may the degree of assimilation.

A number of theories have attempted to account for phoneme recognition. One approach has been to link speech perception with speech production. An early example is Motor Theory, based on the hypothesis that, in processing speech, we relate what we hear to our own experience of forming the same sounds. The theory faces problems in accounting for how we manage to interpret the speech of a speaker with a different regional accent, and how learners of a foreign language succeed in categorising sounds that they are not yet capable of making. A similar link between production and recognition is made in Stevens’ (1960) analysis-by-synthesis model. Here the assumption is that we identify (analyse) phonetic features in the input, then use our competence as speakers to synthesise them.

A second set of proposals suggests that, given its amorphous nature, the phoneme might not be a unit of processing at all. The idea is supported by evidence from illiterate Portuguese subjects who had difficulty in performing phoneme-based tasks; this suggests that phonemic awareness may be a product of literacy rather than vice versa. Some commentators have proposed a larger, more stable unit of processing such as the syllable. The assumption is that listeners match spoken input against a stored set of all the syllables of their language. However, a more efficient unit might be the demi-syllable (in the word cat, /k æ / and / æ t/) which would require fewer forms to be stored (the estimate for French is about 2000).

 Alternatively, a direct match might be made between phonetic features in the signal and the stored representation of words in the mind, without needing to convert the features into phonemes. A phonetic feature detector would examine auditory values over a short time-slice, and would detect the presence or absence of particular features on the basis of the evidence available. This information could then be compared with entries in the lexicon, where the spoken word would be stored in the form of a cluster of features which distinguished it from all others. This is the basis of Stevens’ (1986) LAFF (Lexical Access from Features) model. In an alternative ‘direct’ model, Klatt’s (1986) LAFS (Lexical Access from Spectra), words are stored instead in the form of spectral patterns which represent the transitions between their phonemes.

A third set of solutions focuses upon phonological representation. Because phonemes (and indeed words) are so variable in connected speech, there is obviously no possibility of making an exact one-to-one match between the word form that we hear and a template in our mind. Whatever is stored therefore has to allow for variation. It might be that the phonological form of a word takes the form of an idealised prototype to which a sequence of sounds in the signal can be matched on the basis of an approximate fit. It might be that the stored form of the word is underspecified, consisting only of those features which are sufficient and necessary to identify it. Or it might be that we are capable of storing multiple exemplars of a particular word (all those that we have come across), thus enabling us to form a match between the form we hear and any one of a large number of stored variants.

A final, radical solution to the ‘non-invariance’ problem proposes that we do not package the signal into linguistic units at all but divide it up into equal sections determined by time. This is the approach adopted by the most well-known computer simulation of listening. TRACE, a connectionist model, processes the signal in small time-slices which are independent of phoneme, syllable and word boundaries. Each time-slice is connected to those that immediately precede it, so the processor can combine evidence from current input with evidence from what has immediately gone before. This enables TRACE to deal with the way in which phonemic features overlap in time due to co articulation. However, it faces the problem of adjusting to differences of speech rate between speakers.

See also: Normalisation, Phonological representation, Unit of perception

 Further reading: Hawkins (1999); Nygaard and Pisoni (1995)

اخر الاخبار

اخبار العتبة العباسية المقدسة

مركز تصوير المخطوطات وفهرستها.. جهود منهجية لتصوير التراث وحفظه

قسم التطوير يواصل تقديم دورة أخلاقيات المهنة ضمن البرنامج التأهيلي للمنتسبين الجدد

انطلاق فعاليات الجلسة المسائية لمسابقة النور القرآنية الدولية الجامعية في جامعة العميد

الحزام الأخضر الجنوبي الأول يستضيف وفدًا من محافظة البصرة ضمن نشاطات قسم الشؤون الفكرية والثقافية

المزيد

الآخبار الصحية

أول إرشادات غذائية قائمة على الأدلة لمعالجة الإمساك المزمن

صباح بلا إفطار؟.. صحتك في خطر!

هل يمكن للجلد أن يكشف عن صحتك العقلية؟.. دراسة حديثة تجيب

السر في مطبخك.. 6 أطعمة تمنح عظامك قوة الشباب

المزيد

الآخبار التكنلوجية

ناسا تكشف: كوكب الأرض أصبح أقل لمعانا خلال العقود الماضية

دراسة: الوقت الذي تستخدم فيه هاتفك قد يفضح حالتك النفسية

الهنغاري لاسلو كراسناهوركاي يفوز بجائزة نوبل للآداب 2025

رسميا.. إعلان الفائز بجائزة نوبل للسلام لعام 2025

المزيد

مواضيع ذات صلة

SPREADING ACTIVATION

SPEECH: UNIT OF PRODUCTION

SPEECH SIGNAL

SPEECH PRODUCTION

SPEECH PERCEPTION: AUTONOMOUS VS INTERACTIVE

SPEECH MODE2

SPEECH MODE1

SPEECH CODE

SPECIFIC LANGUAGE IMPAIRMENT (SLI)

SPECIES SPECIFICITY

SPECIAL CIRCUMSTANCES (also EXCEPTIONAL CIRCUMSTANCES)

SPEAKING RATE