Stop Place Contrasts before Liquids Edward Flemming MIT Is there a biological grounding of phonology? • Phonetic/phonological systems are shaped by the need to support rapid, robust communication, • …given the limits of our speech production/perception apparatus. • Biology imposes constraints on phonology. • Efficient communication: Auditory, visual – Maximize the distinctiveness of contrasts systems – Minimize articulatory effort – Maximize the rate of information transmission Vocal tract, motor control,… Restrictions on coronal-lateral clusters • A case study adopting this approach: Stop place contrasts before liquids. • Many languages allow [pl, kl]/[bl, gl] clusters, but exclude initial [tl, dl] (Kawasaki 1982). • E.g. English, German, Norwegian, Thai, etc • English: [b-d-g, p-t-k] contrast before [r], brew-drew-grew, pry-try-cry [b-g, p-k] contrast before [l], blue-glue, plan-clan initial [dl-, tl-] are not possible. Restrictions on coronal-lateral clusters • This is a result of a dispreference for coronal-dorsal stop contrasts before laterals (Flemming 1995). • Some languages reverse the English pattern, [tl-, dl-] OK, but no *[kl-, gl-]. • e.g. Haroi and other Chamic languages (Mudhenk & Goschnick 1977), Katu dialects (Wallace 1969): QuickTime™ and a TIFF (Uncompressed) decompressor are needed to see this picture. • Some languages have free variation between coronal and velar before lateral (but contrast elsewhere), e.g. Bolton English (Shorrocks 1998), Mong Njua (Lyman 1974) QuickTime™ and a TIFF (Uncompressed) decompressor are needed to see this picture. Restrictions on coronal-dorsal contrasts before [l] • Kawasaki (1982) hypothesizes that this dispreference is due to perceptual similarity of [dl-gl], [tl-kl]. • General hypothesis: preferred clusters are those in which all contrasts are perceptually distinct (cf. Ohala 1992, Steriade 1999, Wright 2004, etc). Before [l], contrasts between coronal & dorsal stops are not very distinct. Evidence: • Kawasaki (1982): Evidence from 1 speaker that formant transitions are very similar in [dl-, gl-]. - But bursts can be sufficient to distinguish stops. • Hallé, et al (1998), Hallé & Best (2007): - French listeners identify (illegal) [dl-, tl-] as [gl-, kl-] respectively. - French and Am. English listeners have difficulty discriminating Hebrew [dl-gl] and [tl-kl] contrasts. Restrictions on coronal-dorsal contrasts before [l] • Further investigate Kawasaki’s hypothesis through acoustic analysis of American English and Hebrew. • Would English [dl-gl] and [tl-kl] contrasts be less distinct than stop place contrasts before [r] (and vowels)? • How do we infer the expected realization of [dl-, tl-]? • Other stop-liquid clusters • Medial [-dl-, -tl-] clusters • [tl-, dl-] in languages that allow these clusters (e.g. Hebrew, Russian). • But NB languages that allow these contrasts may realize stop-liquid clusters in a different fashion, e.g. less gestural overlap. Cues to stop place contrasts • Prevocalic stops (e.g. Dorman et al 1977): Release burst - transient + frication Formant transitions burst formant transitions 5000 0 26.1405 26.5479 Time (s) Materials • 6 near-minimal triplets for br-dr-gr • 6 near-minimal pairs for bl-gl • 9 triplets for [b, d, g], each preceding the same set of nine vowels. _r _l _V b brown blow bid d drown did g ground glow gig • Sentence frame ‘Say X to me’ • Presented twice in random order • 5 native speakers of American English, 4 female, 1 male. 5000 5000 0 0.350678 0.47009 0 0.312515 Time (s) bl(ow) 5000 0.487222 Time (s) gl(ow) 5000 0 0.322387 0.522607 Time (s) br(ew) 5000 0 0.337107 0.537411 Time (s) dr(ew) 0 0.352544 0.552427 T ime (s) gr(ew) Quantifying burst shape • Measured from smoothed spectra (Hanson & Stevens 2003) Calculate a series of seven DFTs on 3 ms windows at 1 ms intervals. Average these spectra. Quantifying burst shape • Burst peak: amplitude peak of the burst spectrum bl bursts • Amid-Ahi (cf. Suchato et al mid gl bursts high 2005) •Amid = average amplitude from 1.25 kHz - 3 kHz •Ahi = average amplitude from 3.5 kHz - 8 kHz Measurements Burst: • Burst duration - from stop release to onset of first formant. • Usually coincides with onset of voicing, but weakly voiced frication is included in the burst. Formants: • Measured F2 and F3 at the end of the burst, or the onset of formants. 5000 br 0 58.0934 58.1442 58.1686 58.1755 Time (s) 58.2355 br vs. dr vs. gr - Formants Formant onsets 2500 2000 Hz 1500 * * * 1000 500 0 F2 F3 • Formant transitions distinguish [br, dr, gr]. br dr gr * p < 0.01 br vs. dr vs. gr - burst Amid-Ahi burst duration (ms) * 15 dB 0.05 10 5 frequency of peak * 0.04 br dr * 0.03 0.02 gr 1000 500 0.01 • burst distinguishes [br, dr, gr] * p < 0.05 * 1500 0 0 * 2000 Hz 20 burst duration 0 br dr * gr br dr * gr bl vs. gl - formants Formant onsets 3500 3000 Hz 2500 2000 bl gl 1500 1000 500 0 F2 F3 • Formant transitions do not differentiate [bl] from [gl]. bl vs. gl - burst Amid-Ahi burst duration (ms) * 15 dB 35 10 5 2000 * 30 25 1500 20 Hz 20 15 10 gl • [gl] is more compact than [bl]. * 1000 500 5 0 0 0 bl frequency of peak burst duration bl gl • [gl] has a longer burst than [bl]. * p < 0.01 bl gl • [gl] burst has a higher frequency peak than [bl] burst. Summary • [br, dr, gr] are distinguished by formant transitions and burst quality in much the same way as prevocalic stops. • [bl, gl] are distinguished by burst quality, but not formant onsets. Stop-[l] clusters • Stops are not well differentiated by formant transitions before [l]. • Initial [bl-, gl-] • No significant differences in formant onsets. • Medial [-dl-, -gl-] • Small differences in F2 onsets • -dl- seems to allow more coarticulatory influence of a preceding front vowel. Formant onsets 3500 3500 3000 3000 2500 2500 bl gl 2000 1500 1000 Hz Hz Formant onsets 1500 1000 500 500 0 0 F2 F3 dl gl 2000 F2 F3 Medial [-dl-] vs. [-gl-] • • • • • 3 speakers from the previous study Near-minimal pairs, e.g. wrigley-ridley, badly-Bagley. Same measures as before. Clusters are differentiated primarily by burst quality Small differences in F2, F3 onsets • -dl- seems to allow more coarticulatory influence of a preceding front vowel. frequency of peak Amid-Ahi burst duration 2500 20 2000 25 15 15 10 1500 Hz 20 dB burst duration (ms) 30 10 500 5 5 0 0 0 dl gl 1000 dl gl dl gl Comparison to prevocalic stops Burst spectrum: • [b] - low frequency peak • [bl] is similar. Sometimes additional peak ~3000 Hz (F3 or F4) • [g] generally has a prominent peak < 5000 Hz, frequency dependent on following vowel. • [gl] has a peak at ~1300 Hz, cf. goal d- 40 40 –30 0 –30 0 40 2500 Hz -dl- 2500 Hz 5000 5000 40 goal -gl- –30 –30 0 40 2500 Hz 5000 0 2500 Hz 5000 Summary • Stop place contrasts are distinguished by burst characteristics and formant transitions in stopliquid clusters. • Both types of cues appear to be comparable to prevocalic cues to stop place in most cases. • The perceptual significance and relative weight of these cues remains to be assessed. Restrictions on coronal-lateral clusters • Preliminary study of medial [-dl-,-gl-] in English - difference in formant transitions in some contexts. - difference in burst spectrum (Peak: [dl] 1579 Hz, [gl] 1177 Hz) - but medial [-gl-] is a bit different from initial [gl-] 5000 5000 0 0.276954 5000 0 0.524867 0.278185 Time (s) (ba)dly 0 0.499791 0.000422777 Time (s) (Ba)gley 0.2216 Time (s) glee Restrictions on coronal-lateral clusters • Hebrew contrasts initial [bl, dl, gl], [pl, tl, kl]. 5000 40 dl –30 0 0.140447 5000 0 1000 2000 0.335739 3000 4000 5000 Frequency (Hz) 6000 7000 8000 Time (s) 40 gl –30 0 0 0.120813 0.320904 Time (s) 1000 2000 3000 4000 5000 Frequency (Hz) 6000 7000 8000 Introduction • A case study adopting this approach: Stop place contrasts before liquids. Background: • All languages allow stop place contrasts before vowels. • Many languages permit word-initial stop place contrasts before liquids. E.g. English: [b-d-g] contrast before [r], brew-drew-grew [b-g] contrast before [l], blue-glue • In some languages, stop-liquid clusters are the only onset clusters, e.g. Spanish. Introduction Structure: • Dispreference for dl-gl • Theory: insufficiently distinct. • Evidence: other stop-l clusters, medial, other lgs, Tr clusters for comparison (3-way contrast). • Cues to stop place: burst, formant transitions • Measures • What’s wrong with dl-gl? • Tl clusters not differentiated by formants – English initial, medial, Hebrew initial – compare Tr clusters – Why? coproduction with (velarized) [l] - normal for b, g, not for d • Tl clusters are differentiated by burst durn, quality – but [dl] bursts are non-canonical, shifted towards velars • So 3 way contrast would have to be realized by burst - less Introduction Two questions: • Why are stop-liquid clusters preferred onset clusters? • Why are [tl-kl], [dl-gl] contrasts dispreferred? Hypotheses: • Preferred clusters are those in which contrasts are perceptually distinct. • Stop place is generally well-cued before liquids. • But coarticulatory effects render coronal and velar stops perceptually similar before [l] (Kawasaki 1982). Introduction • The preference for stop-liquid clusters is often attributed to a preference for large sonority rise between consonants in an onset. • Alternative: preferred clusters are those in which all contrasts are perceptually distinct (cf. Ohala 1992, Steriade 1999, Wright 2004, etc) • High sonority of liquids is relevant only because more sonorous sounds are generally better able to support the realization of cues to adjacent consonants. • This line of analysis can be extended to account for restrictions that are unrelated to sonority: dispreference for coronal-velar contrasts before laterals. The present study • Investigate acoustic properties that could serve as cues to stop place contrasts in stop-liquid clusters, using American English as a case study. • Many languages permit stop place contrasts before liquids. E.g. English: [b-d-g] contrast before [r], brew-drew-grew [b-g] contrast before [l], blue-glue initial [dl-] is not possible. • What is the nature of cues to place in this position? How doe they compare to place cues in prevocalic position? • Why are liquids good contexts for stop place contrasts? • Why is the pre-[l] context not as good as the prevocalic context? Specifically, why are [dl-gl]/[tl-kl] contrasts often excluded?