Getting at variation with ultrasound: Scottish and Dutch /r/ Ultrafest 3 University of Arizona at Tucson April 2005 James M Scobbie (QMUC) Koen Sebregts (Utrecht) and thanks also esp. to Alan Wrench and Yolanda Vasquez Alvarez

Why /r/? Why ultrasound? Previous articulatory studies of /r/ –There is a lot of variation Variation, change, acquisition, phonetics –“labiodentalisation” of /r/ in Anglo English –“vocalisation” of final /r/ in Scottish English –“retroflexion” of final /r/ in Dutch Phonology – what *is* an /r/ anyway? Ultrasound –Relatively informal –Can provide dynamic images

Can we add to MRI? Tiede’s beautiful images (from ICPhS 03)

Can we add to MRI? And perhaps what we can expect…

Can we add to MRI? And perhaps what we can expect…

Can we add to MRI? Tiede’s beautiful images (from ICPhS 03)

Problems with ultrasound The usual –Incomplete images, no passive articulator –Head-probe correction or control –Synchronisation and low frame rate –Splines and edges –Stats plus… –European video output (PAL) is at 25Hz (albeit with more pixels, esp in raw AVI) –Fieldwork taped data is slower to prepare

Data collection – the laboratory Good –Helmet (or other head games) –Less environmental noise –The experimenter is in control –Choice of equipment and software - demodemo but… –Willing and normal subjects have to be found –Experiment costs –Intimidation of subjects

Data collection – the “field” In this case – Glasgow Science Centre Good –Lots of varied and willing subjects –Outreach… and publicity (!) –Vernacular speech more elicitable –Qualitative articulatory transcription - demodemo but… –Things are a little out of control (cf feedback) –Small amount of time for each subject –Non-ideal equipment and methodology

What do we want for /r/? Varied vowel environments Varied word and syllabic positions Acoustic analysis (and other channels?) Info on multiple articulators Stratified pool of subjects Normal vernacular speech Dynamics Synchronised data

Backstep: methodological issues Smearing of raw data due to scan rate Creating of video output - downsampling and synchronisation problems

Matrix of raw data generated at the scan rate of the ultrasound machine e.g 72 complete scans per second Matrix of raw data Image processing Buffer of image data Video scanning Buffer of image data Smooth greyscale image buffer i.e. video output What’s the delay?

Methodological issues There is always a delay –Our range was 20ms to 100ms (mean 40ms) –Does our 25Hz rate make it clearer or worse? Practicalities –Multichannel synchonisation, even video, even acoustics, is based on unpredicatable delay whether via camcorder or direct –Individual video frames can be arbitrarily split in addition to overlay and interlacing –High ultrasound sample rate alone is not enough

Scottish background Field work in Glasgow Science Centre Stuart-Smith et al discover heavily retracted coda /r/ in young (teenage) vernacular Scots –Not “vocalisation” like middle class Anglification –Strong breaking (transitioning) with limited time at target? –Different target? –Limited evidence of mergers (yet)

Scottish pilot Methodological –Probability of numerous subjects (hundreds) –All age groups, wide spectrum of social mix –Handheld probe plus mike mix to tape –Eyeball qualitative analysis is highly feasible –Need lab-based follow-up for quantitative research Descriptive so far (very sketchy!) –Clear and obvious cases of pharyngeal /r/ –No meta-linguistic awareness of change

Scottish hypotheses & implications Strong breaking in coda is a strong pharyngeal or tongue dorsal gesture Some speakers have lost any obvious anterior gesture in coda Anterior gestures, if present, include retroflex and bunched types of /r/ This is socially stratified –There should be intra-speaker variation too –How categorical are these variations? We need representative dynamic data

Dutch background Large sociolinguistic, phonetic and phonological survey of Dutch (van Hout, Zonneveld and Van der Velde) –400 subjects in multiple locations in Netherlands Some speakers have uvular trill/fricative /r/ in onset, and an anterior approximant in codas –What is going on articulatorily? –What is the inter-speaker sub-variation like?

Dutch study Subjects –10 – all in Edinburgh –Post-screened down to 4 anterior /r/ users Materials and protocol –Picture naming (n < 30) –Real words, near minimal pairs: /ir/ /ur/ /ar/ –Singleton /r/ and cluster /r/ and /r/-less –3 reps Feedback –Unaware of focus on /r/

Dutch study Target – it’s a multi gestural thing –How do we choose the right frame for /r/? –Using acoustics needs good synchronisation –Using the images themselves is circular –Intergestural timing at 25Hz? Dynamics –How to characterise? Acoustics –Same sort of questions – target & dynamics –Final devoicing…

Boer MS – point is red at end of voice – tip down

Boer 1 RB

Boer 2 RB

Boer 3 RB

Mier 1 RB

Mier 2 RB

Mier 3 RB

Schaar 1 RB

Schaar 2 RB

Schaar 3 RB

Boer 2 RB

boer 2 vdl

boer 2 VDB

boer 2 MS

So far Visual inspection of raw images or dynamic spline diagrams –Two retroflexers (RB VDL) –Two bunchers (MS VDB)

Next Acoustic identification of an r phase –Midpoint spline can be extracted –No discrimination of voiced or voiceless –Identification of single “max rhotic” is similar Acoustic analysis –Steady-state V + transition + r-phase –Location of end of voicing –F2 & F3 of voiced (usually) r target if obvious

MS bunched mainly voiced

VDB bunched mainly voiced

RB retroflex mainly voiceless

VDL… retroflex-ish but nearly vocalised?

VDL… real non-rhotic vowels for comparison boer vs. koe mier vs. riem schaar vs. sla

Dutch summary 4 speakers with anterior coda /r/ –Two retroflex –Two bunch One of the retroflex speakers is gradiently vocalising –Timing the /r/ late before pause –But some long domain cues in vowel quality and consonant variation –RB has strong transitions too rather than steady state but ?? sounds more rhotic Interesting to look at following C…

Overall summary Dutch anterior /r/ has various flavours –Retroflex –Non-retroflex –Weakened and late pre-pausal gesture Glasgow /r/ comes in various flavours –Pharyngeal approximant observed –Breaking taking over from “rhoticity” Strong phonetic effects –Vulnerable non-standard speech –Socially-stratified qualitative UTI is T.o.C.

Conclusions, implications Weakening of final /r/ is flexible –Multiple articulators provide lots of options –Strength and timing are affected –Effects can be gradient and/or categorical –Acoustic effects appear complex –Phonological contrasts need not be affected Small (?) sub-phonemic effects, change, and variation go hand-in-hand Fine detail and structured variation are in the grammar

Methodology – field solution Splines and edges lose a lot of info –Relative to MRI X-ray etc –Relative to ultrasound images Qualitative analysis of images usefully augments impressionistic transcription This can be done live and/or from tapes made in the field Small amounts of data from large numbers of subjects is standard practice

Methodology: lab solution #1 Dedicated hardware + software in the lab Matrix of raw data time stamped at headline sample rate (100Hz) Image processing Audio etc Time stamped PC buffer & files High speed, synchronised, clean images Integrated analysis (spline fit, export etc) Start

Methodology: lab solution #2 Multichannel backbone (incl 200Hz EPG) Funding application submitted –100Hz synchronised machine (+ EPG) –Multichannel developments (EMA, VICON) Helmet improvements Also, funding application submitted for ultrafest 5 (2007)

Holiday report Excellent!

Demo of lab data collection Slide 88 AA