Jonathan Harrington Contextual ambiguities in speech signals and their consequences for sound change.

Slide 1

Jonathan Harrington

Contextual ambiguities in speech signals and their consequences for sound change Historical sound changeSien e sege and e assaut watz sesed at Troye,e bor brittened and brent to bronde and askez,Chaucer, Canterbury Tales, late 14th C.Sincesiegeassaultceasedburgburntashes

People often ask me what made me take up writing... You see I put it all down to the fact that I never had any education.Agatha Christie, 1955

Sound change and language-internal factorsJ. Ohala (1981, 2008) suggests that the conditions for sound change to occur can be met when exceptionally listeners process or parse coarticulation in a way that is different from the way in which it was produced.Parity in coarticulationIn productionA nasalised vowel is heard as ORAL in a nasal context like /m_n/: listeners undo or reverse coarticulation in productionIn perceptione.g. Kawasaki, Beddor and others:

Coarticulation and sound changeAccording to Ohala, some change can come about when listeners exceptionally fail to normalise coarticuation i.e. to normalise for variation in relation to its context.flapThe earth used to be flat but now it's a sphere

Normalisation for context in vision

Insufficient normalisationmanbadIn contextIn context

Insufficient normalisationLatin: 'manus' > French 'main' = /ma/ (Hand) Associated sound change

Others (e.g. Bybee, Lindblom) see the origin of sound in hypoarticulated speech.This is type of speaking style in which the signal is less clear if the words are either predictable from context and/or less important for understanding what is being said.

Hypoarticulated speechCoarticulation and hypoarticulated speech1Perhaps the conditions for sound change are met because the information allowing listeners to process/parse coarticulation is degraded in hypoarticulated speech.To test this in the laboratory, we investigated coarticulation in prosodically accented (analogous to hyperarticulated) and deaccented (hypoarticulated) speech.1. J. Harrington, F. Kleber, J. Siddins, U. Reubold, Laboratory Phonology (2014)The first coarticulatory effect to be investigated using German materials is polysyllabic shortening which is sometimes associated with historical sound changes e.g. south (long) but southerner (short).Polysyllabic shorteningcoarticulationcoarticulationIs the available information to process coarticulation diminished in hypoarticulated speech?Hypoarticulated : MARIA hat sackt verstandenHYPERarticulated : Maria hat SACKT verstanden26 1st language German speakers produced the 4 target words in contexts that were typical of HYPER- and hypoarticulated speech (prosodically accented vs. deaccented)sackt/zakt/sagt/za:kt/V/C (duration) ratio(sags)(says)sackte/zakt/sagte/za:kt/(past tense)

SACKTSAGTSACKTESAGTE

log(V/C)2 syll1 syllHYPERarticulatedshort Vlong V

sacktsagtsacktesagteHypoarticualtedshort Vlong Vlog(V/C)contextCoarticulation(2 vs. 1 syllable)Production: less coarticulatory information in hypoarticulated speech

Is less in hypoarticulated

P(long)Gaussian (logistic) classificationWe tested whether the same subjects as listeners have more difficulty parsing this kind of coarticulatory information in hypoarticulated speech. To do so, we synthesised continua between sackt-sagt and sackte-sagte such that the target word was accented or deaccented by manipulating only the fundamental frequency (not duration!).The listeners task was to label (forced-choice) each stimulus as one of sackt, sagt, sackte, sagtePerception

AnnahatteSACKT(E)-SAGT(E)verstandenANNAsackt(e)-sagt(e)

Influence of context (2 vs. 1 syllable)log(V/C)HYPERarticulatedHypoarticulatedP(long)Context has a bigger influence on decisions in HYPERarticulated.i.e. listeners are less accurate in parsing coarticulation if they are led to believe that the speech is hypoarticulatedPerception(Vertical axis: proportion of long i.e. sagt(e) vs. sackt(e) vowel judgments)Intermediate summaryThe conditions for sound change are met when listeners fail to process (parse) coarticulation in a way that is consistent with its production.This mismatch between the modalities of production and perception is especially likely under conditions in which the signal is degraded due to a hypoarticulated speaker style.Dissimilation: in which a repeated sound is deleted. (e.g. Grassmans law, Ancient Greek /trikhos/ (hair genitive) presumably derived from /thrikhos/ (cf. /thriks/, hair nominative)Sound change and dissimilation1Ohalas model: Listeners incorrectly interpret the first /h/ as due to coarticulation caused by the second/ h/ (i.e. an over-application of context in which coarticulation is incorrectly factored out.)1. Mary Stevens & Jonathan Harrington, Paper presented at 3rd sound change workshop, San FranciscoDissimilation and analogy to vision

1. http://www.welt.de/wissenschaft/article11443233/Darum-fallen-Sie-auf-optische-Taeuschungen-rein.htmlAn over-application of perspective (caused by the room being crooked) causes the illusion that one of the monozygotic twins is bigger than the otherWe investigated the basis of /w/ dissimilation as shown by changes such as:/knkwe//kwnkwe//tikwe/ (cinque)But no /w/ deletion/kwndekim//kwnditi/Italian(fifteen)LatinSound change and dissimilationAn 11-step canto-quanto /kanto- kwanto/ continuum (sing, how much)13 listeners heard this continuum before quattro and sette

F2 was also lowered in anto to simulate lip-roundingSound change and dissimilationIf so, they should hear more canto preceding quattro (because the 2nd /w/ masks the first /w/ in /kwanto kwatro/)1. Ohala: a repeated /w/ can make the first /w/ hard to hear because listeners interpret the first /w/ as due to coarticulation2. (based on results from sackt/sagt). The effect in 1. might disappear when the continuum is deaccented.Predictions detto KWANTO-CANTOThe continuum was ACCENTED or deaccented DETTO kwanto-kantoquattrosettevolteHo

_ quattro_ settequantocanto/kw//w/Sound change and dissimilationresponseacoustics

The word effect disappears for deaccented (consistent with 2)deaccentedaccentedMore quanto responses before _quattro for accented (contra 1)Results3. More canto responses when deaccentedPredictions1. More canto responses in a quattro context (Ohala)2. The influence of coarticulation is less in deaccented (Harrington)Link to sound changeThe identification of the first /kw/ in Latin quinque depended on lip-rounding spread through the word which listeners had difficulty hearing, when quinque was hypoarticulated (leading to a perception of /kinkwe/) .DissimilationNo evidence from these results for Ohalas idea that dissimilation comes about through perceptual maskingInstead: An initial /kw/ benefits from a final /kw/ (of quattro) but only in accented wordsIn deaccented (hypoarticulated) words, this effect disappears and there is a bias towards cantoi.e. the conditions for sound change are met in hypoarticulated speech because in this speaking style, processing coarticulation is more difficult for listeners.SummarySound change and external factorsSound change could also come about because speakers do not process coarticulation in the same way.1st language listeners who vary in the extent of coarticulation in production and its normalisation in perception (Beddor, 2009; Baker et al, 2011).Different speaker groups: e.g. Yu (2013) has shown that normalising for coarticulation is influenced by personality traits such as autism.Dialects/languages have different degrees of coarticulation.Sound change and external factors: adults and childrenSound change may also come about because children normalise less effectively for coarticulation than adults.Some evidence again from visione.g. In contrast to most adults, children under 7 years tend to perceive (correctly) that the circle on the right is larger than the one of the left

Doherty et al, 2010, Developmental Science: Young children are less accurate than adults when context is helpful, but more accurate when context is misleadingThe general hypothesis to be tested is that for children vs. adults:1. The magnitude of coarticulation in production may be similar (or greater)2. Childrens ability to normalise for coarticulation may be lessIf 2, then their relative inability to normalise for context may be a source of sound change.Coarticulation: adults and children11. Felicitas Kleber & Sandra Peters, Presented at Laboratory Phonology, Japan, 2014.Children and coarticulation: present studyMaterials/, / are contrastive in German (musste/msste: had to/ should) and are acoustically differentiated primarily by a lower/higher second formant frequency.Nonwords: /pp, pp, tt, tt/Subjects13 L1-German pre-school children aged between 5-6 years.20 L1-German adults.Labialisation in /p_p/ shifts // towards //Tongue-tip contact in /t_t/ shifts // towards //Coarticulation

Tutt

Ttt

Pupp

PppPictures of real words (e.g. ski, shoe) and of characters whose names were /tt/, /tt/, /pp/, and /pp/ 2 vowels 2 contexts 5 repetitions 13 children = 260 tokens 2 vowels 2 contexts 20 repetitions 20 adults = 1600 tokensProductionChildrenAdultsQuantifying coarticulationWe obtained a relative measure separately per speaker of the distance, d, of any vowel (V) to the means of /, / in /pp, tt/ (in which consonant acoustics reinforce the /, / separation) in an F1 F2 space (formant values extracted at the vowels temporal midpoints).ppttVd = log(d/d)d is negative: closer to ppd is positive: closer to ttd is zero: equidistant between the twoF1F2ddQuantifying coarticulationNearer ttNearer ppd0highCoarticulation in pp lowlowCoarticulation in tt highResults: coarticulation in production1. More coarticulation in /tt/ than in /pp/2. No differences in the magnitude of coarticulation between adults and children.

(The nearer tt, pp are to d = 0, the greater the coarticulation)AdultChildAdultChildPerception (method)We synthesised an 11-step /-/ continuum and embedded it in /p_p/ and /t_t/ contexts.Adults: forced choiced perception test: identification of each stimulus as one of PUPP, PPP, TUTT, TTTChildren: an imitation task they produced one of the above four names in response to each stimulus.Maria hat CVC gesagt

13 children = 858 tokens 20 adults= 1320 tokens 2 continua 11 stimuli 3 repetitions Perception (method): imitations and categorical responseIn order to compare with adult categorisations, we classified each childs imitation as // or // after (Gaussian) training on the same childs speech production tokens (separately by consonantal context).

Xe.g. X = childs imitation of a stimulus from the perceptual p_p continuum/, / are distributions of productions by the same child in the same p_p context in a formant space.X is categorised as /Y/Some predictions1. (Trivially) Listeners normalise for coarticulation in relation to how it is produced i.e. more // perceived in t_t than in p_p ppppProductionDistance ratio

tttt

Perceived as2. Perceptual normalisation: less for children?

AdultsChildrenResults: less perceptual normalisation for context by childrenAcoustic synthetic stimulus//-like//-likePerceived asp_pt_tConclusionsSound change can occur when listeners misinterpret how phonological contrasts are associated with articulatory information that is temporally distributed in the speech signal.This misinterpretation is especially likely in a hypoarticulated speaking style in which (a) the coarticulatory information may be less saliently communicated by the speaker and/or (b) listenerscapacity for processing coarticulation is diminished.Developmental factors may be relevant in particular if childrens ability to resolve coarticulation as listeners is weak in relation to the magnitude of coarticulation in their own speech production.