The role of voice quality - PowerPoint Presentation

1. The role of voice quality in the Polish laryngeal contrast Geoff Schwartz, Maral Asiaee, Kamil Kaźmierski & Ewelina WojtkowiakAdam Mickiewicz University in PoznańAPAP 2023, Lublin2

2. OutlineBackground on laryngeal phonologyPrevious phonological descriptionsPhonetic complexity of voice contrastsOur working hypothesis and phonological interpretation[fortis] is active in the phonology of Polish, [voice] is notPhonation in the word-initial voice contrast in PolishPhonation in Polish listeners’ perception of the (incompletely) neutralized word-final voice contrast3

3. “Voice” contrasts in two-way systemsOften described in terms of VOT (Lisker and Abramson 1964); “aspiration” languages: aspirated /ptk/ and short-lag, unvoiced /bdg/ - e.g. Germanic languages, notably English“true-voice” languages: plain voiceless /ptk/ and pre-voiced /bdg/ - e.g. Slavic languages, such as PolishHow should we represent two-way laryngeal contrasts in obstruents?Different theories propose different solutions;How important is the actual phonetic implementation of voicing?That depends on the approach4

4. Binary approachThe “traditional” view (Beckman et al. 2013); Binary feature values: [+voice] and [-voice]; phonetic implementation of contrast not important for the underlying representation.Important implications: Aspirated and plain voiceless stops both described in terms of [-voice]Pre-voiced and unvoiced stops both described in terms of [+voice][pʰ] and [p] [b] and [b̥] phonologically equivalent5

5. Unary approachDirect encoding of the VOT typologyOften referred to as “laryngeal realism” – a name coined by Patrick Honeybone but the ideas were not so new:LanguageSeriesLombardi (1991)Harris (1994)Honeybone (2005)English/p, t, k/[aspiration]{H}|spread| /b, d, ɡ/ØØØPolish/p, t, k/ØØØ /b, d, ɡ/[voice]{L}|voice|6

6. Unary approachIssues with “Laryngeal Realism”:ignores laryngeal cues other than VOT (e.g. f0, F1 transition), which tend to show parallel behavior in voicing and aspiration systems (Kirby & Ladd 2016; Schwartz et. 2019);Languages like Swedish with both pre-voicing and aspiration (how to motivate overspecification); Does not predict active [-voice] in true-voice languages, but see e.g.;Polish: Rubach (1996), Cyran (2014) for sandhi-voicing dialectsHungarian: Blaho (2008)Breton: Iosad (2012)Dutch: van der Hulst (2015)Moro: Bennet and Rose (2017)Bakairi, Japanese, Azerbaijani, Western Bade (Őri 2023) 7

7. The complexity of laryngeal contrastsHere is a collection of cues that are relevant for two-series contrastsVOTduration of burst/fricationamplitude of burst/fricationduration of preceding vowelduration of following vowelf0 (pitch) at vowel onset/offsetF1 transitionThe weight of these cues can vary systematically across languagesVOT is a more reliable cue in English than Polish (Keating 1980)French: f0 is more reliable than F1 transition; Polish: F1 more reliable than f08

8. The complexity of laryngeal contrastsThe complexity of two-series laryngeal contrasts complicates accounts using [voice], which suggests a single phonetic propertyRather, a wide range of phonetic cues can signal the presence of a more abstract feature, let’s call it [fortis]English: the connection between the laryngeal contrast and voice quality is fairly well-documentedglottal reinforcement/replacement of [fortis]  creaky phonation some evidence for breathiness induced by lenis codas (Sanker 2019)what about in true voice languages such as Polish?9

9. The present studyHypothesis: voice quality is a correlate of the voicing contrast in PolishFortis consonants should induce stiffer phonation on neighboring vowelsListeners should be sensitive to these effectsWe present data from two experiments on Polishone on the production of the laryngeal contrast in word-initial positionthe other on the perception of the (incompletely) neutralized contrast in word-final position    10

10. Study 1 11

11. Production of initial voicing contrast in PolishResearch Question: does the voicing of initial consonants have an effect on the voice quality in the following vowel?In other words, is the vowel in pas produced with a stiffer voice quality than the vowel in bas?12

12. Methods15 monolingual Polish speakersAll females, aged 17-38 (median age: 25);The dataset: 48 Polish plosive-initial words, counterbalanced for voicing (24 voiced, 24 voiceless) and POA (16 labial, 16 coronal, 16 dorsal);Followed by a non-high vowelMono- or di-syllabic Stimuli presented using PowerPoint slides;Participants recorded directly onto laptop, using a USB interface and a head-mounted mic in a quiet room13

13. Methods (cont)Annotated by hand in Praat (Boersma and Weenink 2022)Measurements extracted using Praat scripts:Script 1: VOT durationScript 2: F1 (in Bark difference; F1-f0) and f0 (in Hz) from the first 20% of the vowelAcoustic measures of voice quality using VoiceSauce (Shue et al 2011)Data frames every 1 ms throughout vowel following onset consonantSpectral tilt measures: H1*-H2*, H2*-H4*, H4*–H2kHz*, H2kHz*–H5kHzStatistical analysis: R version 4.0.4 (R Core Team 2021)Mann-Whitney U test was performed to find out whether spectral tilt values in vowels differ after voiced as opposed to voiceless obstruents14

14. Methods (cont) – Spectral tiltSpectral tilt refers to the relative amplitude of different harmonics in the vocal waveH1*-H2* - amplitude of 1st harmonic minus amplitude of 2nd H2*H4* - amplitude of 2nd minus amplitude of 4th harmonicH4*H2K* - amplitude of 2nd minus amplitude of harmonic nearest 2000 Hz. H2K*-H5k – amplitude of harmonic nearest 2000Hz – harmonic nearest 5000 HzAsterisks mean measures corrected for vowel qualityIn general, larger differences associated with slacker phonation, smaller (and negative) differences for stiffer phonation15

15. Methods (cont) – Spectral tiltbas (left) and pas (right) – spectrum of vowel16

16. Results – ‘traditional’ cuesAs far as the parameters traditionally associated with laryngeal contrasts: VOT: F1(F1-f0 in Bark) & f0 (Hz)VOT: voiceless stops slightly aspirated; pre-voicing very commonThe effects of underlying voicing were detectable in both the f0 (p=.019) and F1 (p=.003) parameters SeriesVOT (ms)Std Deviationvoiceless41.70ms16voiced-92.05ms26ParameterOnsetAverageF1 (F1-f0; Bark)voiceless4.9 Barkvoiced4.4 Barkf0 (Hz)voiceless195 Hzvoiced186.6 Hz 17

17. Results – voice qualityVoice quality results: small but significant effects in all four measuresStiffer phonation for [fortis] in green; Less stiff in yellow18Spectral tilt measuresVoicelessVoicedH1*-H2*6.22 (9.93)5.98 (10.6)H2*-H4*3.14 (15.1)5.68 (16.8)H4*–H2kHz*9.60 (15.4)10.8 (12.5)H2kHz*–H5kHz19.8 (23.8)15.8 (20.4)

18. Results – Interim summaryHypothesis partially supported, depending on the measure you look atMore research is needed to determine the relative weight of individual voice quality measures for Polish19

19. Study 220

20. IntroductionSchwartz et al. 2018, 2021: production study of final devoicing using nonce words and and a purely auditory singular formation taskDeals with orthography issue in studying final devoicing (cf. Roettger et al. 2014)speaker hears: Szeby/szepy występują w Warszawiespeaker says: A w Poznaniu jest tylko jeden szeb/szepEvidence for incomplete neutralization (IN), answering common critiques of IN studiesSchwartz et al. 2018 Perception test using stimuli from production studyListener hears: szeb/szepListener chooses between two plurals: szeby/szepyStimuli with normalized vowel and closure durations79 listeners, over 80% accuracy rate !! (OMG)89% accuracy for underlying voiceless, 72% for underlying voiced21

21. MethodsRepeat, over 80% accuracy rate; and the standard cues for final voice contrasts (vowel and closure duration) were controlled for!We were surprised, so we performed more detailed acoustic analysis on the stimuli, including . . .f0 (pitch)burst durationF1 at vowel offsetAcoustic meaures of voice qualityVoice quality measures performed in VoiceSauce22

22. Methods – statistical analysisFirst, the acoustic differences between voiced-voiceless pairs in the stimuli were calculated, the magnitudes of those differences were plugged in as predictors of accuracypositive difference (in expected direction; e.g. pitch higher for underlyingly voiceless consonants)pitch, F1, burst duration, plus 6 voice quality measuresThen a Principal Components Analysis (PCA) was performed in R to try to identify which acoustic parametersA significant interaction found between Component 3 and underlying voicingNegative values of Component 3 increased listener accuracy for underlying voiceless consonants; no effect for underlying voiced23

23. Resultsinteracation with underlying voicing24

24. Resultsmeasures in blue: improved accuracy;measures in red: hindered accuracy25

25. Results – summary of PCA analysisf0 (pitch) and two voice quality measures (CPP; H1*-H2*) did not contribute to accuracyFour voice quality measures (H2*-H4*, H4*-H2K*, H1*-A2, H1*-A3) as well as burst duration and F1 did contribute to accuracyH4*-H2K* had the largest effect, followed by H1*-A3 (A3: harmonic nearest the third formant)26

26. Summary of the results of both studiesVoice quality on neighboring vowels is an additional phonetic correlate of the voice contrast in Polish consonants;H2*-H4* and H4*-H2K*Hz played a role in both the initial and final contrast;H1*-H2*, however, did not play a role in either positionA question for further researchH1*-H2* has well-documented effects in languages with ‘phonological’ phonation contrasts, but apparently not in PolishCould it be the case that in a language like Polish H1*-H2* plays more of a role for speaker identification, while the other measures have more weight for the laryngeal contrast?Voice quality research is challenging, but exciting 27

27. Implications for two-series laryngeal contrastMore research is needed, but it appears voice quality is a phonetic ingredient in two-series laryngeal contrasts, even in so-called ‘voicing’ languages like PolishThis is all compatible with a growing body of evidence that in voicing languages, a [fortis] feature is active, but a feature [voice] does not have to beBlaho (2008), Iosad (2012), Cyran (2014), van der Hulst (2015), Őri (2023)Stiffer voice quality appears to be an additional phonetic cue to this feature [fortis], along with things like pitch and F1 transitionsLaryngeal features are substance-free, which is why the contrasts may be so phonetically complex28

28. Thank you for listening Email us for references:geoff@amu.edu.pl, marasi@amu.edu.pl, kamil.kazmierski@amu.edu.pl, ewelina.wojtkowiak@amu.edu.pl 29