A replicable acoustic measure of lenition and the nature of variability in Gurindji stops

1.1.1. Phoneme inventory

Gurindji’s phonological inventory is typical of many Pama-Nyungan languages, comprising a five-way place of articulation distinction for obstruents and corresponding nasals, three laterals, three glides, and a tap/trill, shown in Table 1. Gurindji makes no contrasts in terms of voicing, consonant length, or frication, and accordingly obstruents are transcribed in Table 1 using the conventional voiceless IPA symbols. Phonetically, the pre-palatal obstruent /c/ is realized consistently as an affricate by the speaker we study (Ennever, 2014b) and so is excluded from the present study. Like many Australian languages, the vowel system of Gurindji is sparse, contrasting three qualities and length (Meakins et al., 2013), shown in Table 2.

1.1.2. Morphological and prosodic structure

Primary stress falls on the initial syllable of Gurindji words without exception. The stress system has not been studied in detail, though broadly speaking, it resembles those of many Pama-Nyungan languages, with secondary stress on most suffix-initial syllables and alternating stress otherwise (Dixon, 2002, p. 557). A consequence is that word-initial syllables fall at the left boundaries of both the morphosyntactic word and a prosodic word. In this study, we examine intervocalic stop phonemes in word initial position (i.e., flanked on the left by the final vowel of a preceding word), and in morpheme-medial position. These two positions contrast in terms of (non-)adjacency to both morphosyntactic and phonological word boundaries. Given the state of knowledge of Gurindji’s stress system, we make no specific claims about foot boundaries, other than to note that word-initial tokens will always be foot-initial also.

1.2.1. Synchronic allophony

Allophonically, stops in Australian languages are commonly reported to possess lenited variants when flanked by vowels and/or liquids (Dixon, 2002; Evans, 1995). Non-coronal and palatal stops may possess corresponding glide allophones, and alveolar stops flapped or tapped allophones. Fricative allophones are less common but have been reported in similar environments (Fletcher & Butcher, 2015; Dixon, 2002). In terms of positional factors, word initial lenition is generally dispreferred, although some Australian languages have lenited allophones in word-initial position (Blevins, 2001). Lenition has been correlated with stress in Murrinh Patha (Mansfield, 2015) and Yir Yoront (Alpher, 1988). Most reports of allophony are impressionistic; however, Ingram et al. (2008) investigate spectrographic data to identify a range of connected speech processes involving reduction in Warlpiri, a Ngumpin-Yapa language related to Gurindji. These include: Stop voicing, trilling, nasal weakening, vocalization, deletion, nasal-stop cluster reduction, and labialization. Other than Ingram et al. (2008), much of the instrumental phonetic work conducted on Australian languages has focused either on place of articulation (Bundgaard-Nielsen et al., 2012, 2015; Butcher, 1995; Tabain, 2009; Tabain & Butcher, 2015) or on those few languages that contrast two series of stops (Butcher, 2004; B. Evans & Merlan, 2004; McKay, 1980; Stoakes et al., 2007). Here we address the resulting gap in our understanding of Australian languages, with respect to manner of articulation.

1.3.1. Duration

One of the most commonly cited factors affecting lenition is rate of speech and segmental duration (Donegan & Stampe, 1979; Gurevich, 2008; Lindblom, 1983, 1990; Shockey & Gibbon, 1993; Zwicky, 1972). Kirchner summarizes the relationship (2001, pp. 217–218):

“…fast speech, by definition, involves shortening of articulatory gestures. This shortening can mean one of two things: either the articulator reaches the target constriction faster, or the constriction itself is shorter.”

It is under these conditions that we also expect articulatory undershoot resulting in acoustic lenition. Soler and Romero (1999), for example, find duration and degree of constriction to be highly and positively correlated in Spanish spirantization phenomena. In the Scouse variety of English, Marotta and Barth (2005) find fricative and approximant allophones to be successively shorter than their stop counterparts. Furthermore, the relationship is understood to be gradient rather than categorical. In American English, stop lenition is reported to be increasingly frequent and pronounced at successively quicker speech rates, and in successively less formal registers (Warner & Tucker, 2011). Kirchner (2001, p. 4) proposes an implicational hierarchy to this effect, claiming that “if a consonant lenites in some context, at a given rate or register of speech, it also lenites in that context at all faster rates or more casual registers of speech.” Taken together, these studies would suggest that, ceteris paribus, the shorter the duration afforded to a constriction, the less likely full constriction will be achieved.

1.3.2. Place of articulation

Place of articulation of the target segment has also been suggested to affect lenition. Foley (1977), for example, proposes a strength hierarchy of places of articulation ordered by their likelihood of undergoing lenition: Velar > bilabial > alveolar. Evidence supporting this is generally constrained to studies of the Romance languages—for example Florentine Italian (Dalcher, 2006) and Balearic Catalan (Wheeler, 2005, pp. 320–324). Divergent patterns are reported in many of the worlds languages (see Kaplan, 2010 for a summary). Explanations for differences in lenition rates based on place of articulation have been couched in terms of physiological and aerodynamic factors (see Lavoie, 2001, pp. 133–138 for velars; Hualde & Nadeu, 2011 for bilabials).

Within Australia, evidence for place of articulation effects is typically marshaled from the extensive reconstruction of diachronic sound changes. One of the most striking sound changes affecting a number of Australian languages is word initial weakening and the loss of stops consonants—a process affecting bilabial, laminal, and velar obstruents but to the exclusion of apicals (Blevins, 2001; Koch, 2004). An additional set of well established historical changes concern languages that formerly possessed a two-way stop contrast. In a subset of these cases we find that the obstruent system has reduced to a single stop series for all places of articulation except for the apicals where a stop contrast is maintained (as for example in some dialects of the Yolngu languages) (Wood, 1978). The accepted path for this phonological re-organization is an intermediate stage of stop-glide lenition affecting the lenis peripheral and laminal stops (Dixon, 2002). Similarly, in the synchronic domain, Mansfield’s (2015) sociophonetic study of lenition in Murrinh Patha notes that peripheral stops are more prone to lenite to approximants than coronal stops. Finally, cross-linguistic surveys of morphophonological alternations similarly demonstrate that peripheral and pre-palatal obstruents undergo lenition more frequently than their apical counterparts (Round, 2010).

Nevertheless, there is also synchronic and diachronic evidence for apical lenition. Taps are found as allophones of apical stop phonemes in a number of languages (see Dixon, 2002) and have been implicated in an intermediate stage of stop allophony preceding the emergence of three rhotic phoneme systems in the Karnic languages inter alia (Breen, 1997; Dixon, 2002). Despite alternations between stops and taps seemingly constituting lenition (i.e., shorter and less complete constrictions), there are no studies closely examining the acoustic properties of taps in Australian languages. Outside of Australia it has been noted that realizations of intervocalic voiced stops, typically transcribed as ‘taps,’ may include some formant structure— a feature more commonly associated with approximants (as reported in American English [Warner & Tucker, 2011]). Since taps have only been impressionistically noted in Australian languages, it is possible that the degree of apical lenition has been understated.

1.3.3. Flanking vowel quality

The present study focuses on the realization of phonemic stops in intervocalic position, widely accepted as the segmental environment most favourable for consonantal lenition (Kirchner, 2001; Lass, 1984, p. 182).¹ There is, however, ongoing research into whether the quality of the flanking vowels themselves has a significant impact on lenition outcomes. Within effort-based models (e.g., Kirchner, 2001, 2004), vocalic openness (or height) is argued to influence lenition rates due to the greater tongue displacement required to make oral closure. Perceptual-based models (e.g., Kingston, 2008) instead contend that consonantal lenition is not sensitive to vocalic openness. Within a perceptual approach, speakers are understood to attend to disparities in intensity between an affected (lenited) segment and its neighbors. In this view, lenition is motivated by a constraint against abrupt interruptions to the intensity contour of a particular prosodic unit, such as those created by a fully occluded stop. Since the intensity differences between consonants and vowels are much larger than the intensity differences between individual vowel qualities, it is argued that consonantal openness is a significant factor in motivating lenition but vocalic openness is not.

Empirical evidence on this issue however is scarce and, as of yet, inconclusive. Competing evidence is found in studies of Spanish lenition alone: Simonet et al. (2012) find less constricted realizations of /d/ after lower vowels than after high vowels, while Colet et al. (1999) and Ortega-Llebaria (2004) find more constricted realizations of /g/ between low vowels. Straightforward expectations arising from claims of articulatory effort are further complicated by the possibility of the consonant in question shifting its place of articulation to co-articulate with the flanking vowels—or vice versa (cf. Carrasco et al., 2012, p. 169). Saltzman and Munhall (1989) find that in cases where there are competing constraints on articulators between vowels and consonants (e.g., [g] in environments /aga/ and /igi/), the location but not degree of constriction for the consonant will vary as a function of the overlapping vowel. There are even fewer studies of Australian languages that have investigated effects of flanking vowel quality on lenition outcomes. Mansfield (2015) reports that following vowel quality was not statistically significant in his study of /p/ and /k/ lenition in the Australian language Murrinh Patha once lexical item was included as a random effect.

We therefore include preceding and following vocalic environments in the current study to probe if there are any significant differences in lenition outcomes on the basis of articulatory effort. We group the vowels based on their proximity to the target consonant’s constriction location. This differs from studies that split vocalic environment into ‘open’ and ‘non-open’ vowels. Instead we anticipate some effort reduction and therefore less lenition for /p/ and /k/ in the environment of /u/ since the former involves lip rounding and the latter involves tongue backing, both of which are articulatory features shared with /u/. In the case of /t/ and /ʈ/ we cautiously anticipate greater co-articulation (and less lenition) in the environment of /i/ due to tongue tip raising, in contrast with /a/ and /u/.

1.3.4. Domain position effects

One final relevant factor affecting lenition outcomes is the position of the target segment within relevant domains. Escure (1977, p. 58) proposes an implicational hierarchy of positions in which lenition operates. She observes that initial lenition is generally less frequent than non-initial lenition at the level of the syllable, word, and utterance. The proposed hierarchy claims that if a language exhibits lenition domain-initially, it will also exhibit lenition in all other non-initial environments. While Escure’s implicational hierarchy has been shown to be violated by a number of languages (see Bauer, 2008), its basic proposal of a dispreference for domain initial lenition has been widely borne out by cross-linguistic surveys (cf. Ségéral & Scheer, 2008). One explanation advanced for this is the importance of preserving phonological information in word onsets, which have been shown to contain acoustic cues critical to word-perception (Marslen-Wilson & Zwitserlood, 1989).

It is also the case that position affects duration (Oller, 1973; Edwards et al., 1991; Tabain, 2003; Cho, 2006), which in turn affects lenition (Section 1.3.1). Consequently we will be interested in this study to probe whether the contributions to lenition of duration and position are to some degree independent.

Finally, usage-based models (e.g., Bybee, Pierrehumbert) predict that tokens in high frequency lexical items are more prone to lenite than tokens in low frequency lexical items.² Such a prediction has been borne out by several lenition studies (Bybee, 2002; Pierrehumbert, 2001; Dalcher, 2006) and so lexical item is included in the present study as a random effect.

1.3.5. The abstract representation of segments

The concrete articulation of a phonetic segment can be regarded as an execution of a more abstract motor plan and/or phonological representation. Theories like Articulatory Phonology (Browman & Goldstein, 1989) propose that such plans contain articulatory targets that may or may not be physically reached given other constraints such as segment duration. Specifically, sequential gestural units can be subject to effects of ‘intergestural sliding’ (Saltzman & Munhall, 1989). That is, when speech rate increases, articulatory gestures tend to ‘slide into each other,’ increasing their temporal overlap, and resulting in the truncation of one or both adjacent gestures. Such processes are typically assumed to be governed by point-attractor dynamics: Articulatory trajectories for a given gestural unit converge on a single state over time, i.e., a single specified target (Saltzman & Munhall, 1989). If this were the case, we would expect any failure to reach the specified target to be the result of duress, such as applied by temporal reduction. On the other hand, if articulatory trajectories need not converge on a single, point-like gestural target but rather a window-like range, there would be grounds for speakers freely producing a range of articulatory velocities and constriction degrees, at least partially independent of temporal reduction.

Parrell (2011) examines Spanish /b/, which like Gurindji stop phonemes, has many unoccluded, sonorous phonetic realizations. Parrell argues that a single, fully occluded articulatory target is sufficient to account for the variation in Spanish /b/, with other realizations the result of articulatory undershoot due to short duration. Parrell also observes that if Spanish /b/ had only an unoccluded target, then one would not expect occluded variants, even under conditions of long duration, yet long, occluded stops are precisely what are found. Like Spanish /b/, the stop phonemes of Gurindji are sometimes fully occluded, thus we have no reason to believe they are represented or planned solely with unoccluded targets. However, will we ask the question, of whether a single, fully occluded target is sufficient to account for the Gurindji data, or whether it is more consistent with there being a range of targets (or the target itself being represented as a range rather than a point), which span full occlusion through to more open articulations.

To be able to answer these kinds of questions acoustically, it is necessary for studies to be able to quantify gradient acoustic variation (such as that involved in stop lenition) and query the extent to which, and the circumstances in which, speakers may diverge from a kinematic system that assumes a point-like articulatory target and set temporal constraints.

1.4.1. Challenges of acoustic speech segmentation

Notwithstanding the advantages just mentioned of acoustic, casual speech data, its analysis presents well-known challenges. The segmentation of continuous speech into discrete acoustic or phonetic units is a somewhat artificial task (Turk & Sugahara, 2006). Ladefoged (2003, p. 103) cautions that “many segments [simply] don’t have clear beginnings and ends” and Fry (1979, p.117) goes so far as to declare that “[from the acoustic point of view] there are only sounds which are more like, and sounds which are less like the vowels of voiced speech.” Concretely, the segmentation of speech sounds presents three challenges: (i) Discretization, (ii) commensurability, and (iii) reproducibility. By ‘discretization,’ we mean the challenge of delineating the edges, by whatever means, of speech sounds. Many speech sounds, whether viewed acoustically or articulatorily, have no point-like onset and offset events, and consequently various proxies are resorted to (Fant, 1973; Lavoie, 2001). Table 3 presents criteria employed for segmenting regular ‘oral stops’ in some recent studies of stop lenition.

By ‘commensurability’ we refer to the challenge of comparing across different segment types. For example, if one uses ‘bursts’ to define the right edge of a true phonetic stop, how should this be compared to the right edge of allophonic variants such as taps (Connell, 1991), fricated stops (Dalcher, 2006), or simple approximants? In Gurindji, this is a pertinent challenge, as a pilot study (Ennever, 2014a) indicates that fewer than 60% of intervocalic stop phonemes’ realizations are true stops, with the proportion dropping as low as 19% for /k/, depending on its position. By ‘reproducibility’ we refer to the challenge of reproducing another study’s results. In practice, due to the challenges of discretization and commensurability, transcription teams may invest significant resources in securing inter-coder reliability, yet in doing so, can converge upon criteria and conventions that differ form those devised in another lab. Moreover, standard instruments have their limits. Consider the stops displayed in Figure 1. The first appears to have a ‘break’ in F2 and F3 (cf. the analysis criteria listed in Table 3) while the second does not, yet Figures 1a, b depict the same token, visualized with different settings of spectrogram parameters—specifically, dynamic ranges of 30dB and 45dB respectively. Because spectrograms paint all intensities as white below some threshold, they can represent regions to the human eye as being ‘empty’ and uniform when in reality they are not, thus distorting the underlying data and inviting false comparison and analysis.

Consequently, a major contribution of this paper is methodological. In Section 3 we introduce a new method for delineating stop-like and approximant-like segments in a manner which addresses our three challenges. It uses the time-varying profile of intensity in certain frequency bands as a basis for discretizing the speech signal in terms of commensurable events (namely, threshold points in intensity velocity functions) in a fashion which is reproducible because it is automated, and deterministic given the acoustic data. Having delineated stop phonemes in this manner, we then measure the change of intensity (Δi) inside the segment, the peak intensity velocity (P_i) and the segment’s duration (D_i), each as reproducible measures of lenition and related quantities.

In previous research, measures of change of intensity (Δi) during a consonant have been employed as quantitative indexes of lenition in studies of Florentine Italian, Spanish, and American English (Bouavichith & Davidson, 2013; Colantoni & Marinescu, 2010; Dalcher, 2006; Lavoie, 2001; Lewis, 2001). Kingston (2008) and Hualde et al. (2011) in particular employ measures of peak intensity velocity (P_i) as a measure of lenition, on the grounds that more lenis variants have less abrupt acoustic transitions, making it difficult to demarcate their edges and hence determine where to measure Δi from. Thus the current study advances a line of research that infers information about lenition from careful measures of acoustic intensity. The novelty of our contribution is to couple this approach with a reproducible method for segment delineation, including of lenited variants and in a manner commensurable with the delineation of fully occluded stops, and to provide explicit arguments supporting the theoretical and empirical validity of the approach.

3.8.1. Linear correlations

As our first test, we measure the linear correlation of Δi versus D_i.P_i. High conformity would support our premise; low conformity would contradict it. Table 7 shows the linear correlation (Pearson’s r) of Δi and D_i.P_i, for our nine frequency bands and four values of the cubic spline smoothing parameter spar. Higher correlation values indicate a closer conformity to (2), and thus by hypothesis, a closer nexus between intensity and articulation.

We interpret these results as follows. Broadly speaking, the greater the degree of smoothing applied to the underlying time series {i(t1), i(t2) … i(tn)}, the more the resulting continuous function i(t) and its derived measures D_i, Δi, and P_i, come to conform to equation (2). Interpreting this cautiously, this may arise because smoothing removes noise which otherwise obscures genuine similarities between intensity and articulation, but it may also be that smoothing reduces jerk and so coerces the data towards a function i(t), whose derived measures D_i, Δi, and P_i, happen to have the properties in (2), or there may be an element of both. However, it can be observed that not all frequency bands are alike. Both the ‘voicing’ and ‘noise’ bands conform less well to equation (2) than the ‘lower’ and ‘upper’ band. There is no reason from the mathematics of spline fitting why this should be so, whereas the observation fits with our predictions, reasoned on phonetic grounds, regarding which bands should more closely mirror articulation. This suggests to us that spectral energy in ‘lower’ bands is a good choice of proxy for degree of constriction, and hence articulation.

3.8.2. Regression testing

In Section 3.8.1 we examined the relationships solely between D_i, Δi, and P_i. Here we apply a more exacting test, asking also how place of articulation, neighboring vowel, position (word-internal versus -medial), and carrier word might affect that relationship. We do this by means of a linear mixed-effects regression model, with carrier word as a random effect. To be clear about what we are attempting to do here: The equation in (2) has just two degrees of freedom, so that if one specifies D_i, and P_i then Δi should be fully predicted. Thus, if our acoustic measures conform to equation (2), we expect that in our regression model D_i, and P_i will overwhelmingly account for the variation in Δi. If additional contributions come from the other factors, even if statistically significant, we expect their effects to be small in magnitude. If that is the case, it offers more reason to believe that our acoustic method is closely mirroring articulation.

Our regression model is summarized in Table 8; variables are explained below. Note that in order to keep the key terms additive, we use the equivalent of the logarithm of equation (2), ln(Δi) = k′_i + ln(D_i) + ln(P_i), where k′_I becomes an intercept term.

The variable Phoneme has three levels. As noted in Section 2.1.1 /t/ and /ʈ/ are pooled word medially as /T/; in initial position /ʈ/ does not occur. For vocalic environment, the dataset was not sufficiently large to test each preceding vowel /i,a,u/ in combination with each possible following vowel (3 × 3 = 9 conditions). Instead, for each stop phoneme we binarily coded the vowel system into vowels that were/were not articulatorily proximal with each stop phoneme as per Table 9 (see Section 1.3.3 for discussion). The resulting binary true/false values for each stop-vowel combination are provided below.

Token counts for each stop phoneme, in environments neighboring true/false proximal vowels to the left and the right, are shown in Table 10.

In total, 581 segments were delimited successfully (out of 586 which had been manually marked-up; see Section 2.2). Speaker was not added as a random effect because the data comes from only 1 speaker. We used a simple additive model because there were not enough data points to test interactions. The model was run using lmerTest (Kuznetsova, 2016) to test for significant predictors and MuMIn (Bartoń, 2016) to provide an R²_C value for the model.⁹ Results are presented in Table 11.

Table 11

Summary of linear mixed effects model. REML criterion at convergence: –1456.3.

As predicted, the model explains close to 100% of the variation in Δi (R²_C = 0.98). It shows that the longer the duration of the stop, the greater the change of intensity Δi and hence the less likely it is to be lenited (p < 0.001), and similarly, the higher the peak velocity, the greater the change of intensity Δi and hence, the less likely a stop will be lenited (p < 0.001). The model suggests some effects of phoneme type, i.e., /p/ does not lenite to the same degree as /T/ (p < 0.05) and some effects for environment, i.e., a stop is more likely to be lenited when it occurs word medially (p < 0.05) and after a proximal vowel (p < 0.01). However, as predicted, the effect sizes of each of these contributions are very small when compared to the contributions of peak velocity and duration. Although they are statistically significant, they barely contribute to accounting for Δi.

In sum, the regression analysis confirms expectations about our acoustic measures. They are behaving algebraically like the articulatory properties they are supposed to mimic. As discussed earlier, there is no inherent reason for them to do that, unless they are tracking articulation closely.