On the salience of prenuclear accents: evidence from an imitation study

2.1 Experimental materials

In order to have controlled variation in PPA patterning, we restrict our study to PPAs in English words that have two possible landing sites for a pitch accent, presented in non-final position of a complex noun phrase in a complete sentence. In English, words where the syllable with primary stress is preceded by another syllable with secondary stress (2–1 words, e.g., èlevátion, òptimístic) allow a less frequent pronunciation with reversal of prominence (e.g., èlevátion → élevàtion), in which case their pattern becomes identical to that of 1–2 words, which have only one prominence pattern (e.g., élevàtor, súpermàrket—which crucially cannot be pronounced *èlevátor, *sùpermárket). This perceptually salient stress reversal phenomenon manifests in production through a High-tone pitch accent, H*or L+H* in the ToBI annotation system (see Beckman et al. 2005), associated with the syllable that bears secondary stress in the unmarked pronunciation of the word. We refer to this accent pattern as “Early High.”

This phenomenon of stress reversal in 2–1 words has been described as particularly common when there is a following accented word within the phrase, where it would be motivated by rhythmic reasons; e.g., Mìssissíppi → Míssissippi Ríver; thirtéen → thírteen mén; Chinése → Chínese lánguage (see e.g., Liberman and Prince 1977, among others). Thus, Early High in lexical 2–1 words has been described as a strategy for resolving stress clash (Horne 1990; Nespor and Vogel 1989), where locating a pitch accent on the initial syllable of a 2–1 word increases its distance from a pitch accent in the following word. Other researchers argue that the Early High pattern arises not by retracting stress from the later syllable, but rather from the placement of an additional pitch accent on the initial syllable (Ross et al. 1992; Shattuck-Hufnagel 1988, 1992, 1995). Grabe and Warren (1995) conclude that stress shift (Early High) is mostly a perceptual phenomenon, driven by listeners’ expectation of rhythmic stress alternation.

We conducted an imitation experiment using constructed sentence stimuli where words with a lexical 2–1 stress pattern were produced in the prenuclear position of a prosodic phrase, with one of three accentual patterns: accent on the lexically specified primary stress (2–1), accent on the initial syllable in the Early High pattern (1–2), and unaccented (but retaining the lexically specified 2–1 stress pattern). By varying the presence/absence and location of the PPA within target words, we could compare the imitation of different, non-contrastive accent patterns on the same word, in the same location within the sentence and within the prosodic phrase. Twelve words with a canonical 2–1 stress pattern were selected as the target words for analysis. All target words had a secondary stress on the initial syllable and primary stress on the third syllable from the beginning of the word, e.g., rèalístic. The complete list of target words is given in (2):

For each of the target words in (2), we constructed a sentence where the target word was the first content word in a noun phrase in sentence initial position (target phrase). The complete list of experimental sentences is given in (3). The phrase containing the target word is in brackets for clarity:^[5]

An adult female native speaker of American English (one of the authors) recorded each of these twelve sentences three times, each time using a different prosodic pattern on the target word, as described in Table 1 and illustrated with the target phrase in (3a).

These three patterns are illustrated in Figure 1. Notice that the second content word in the target phrase was always accented, although its peak is downstepped when the preceding word also carries an accent, i.e., in the Primary and Early High patterns. Notice also the presence of a pitch rise on the final syllable of the head noun, marking a prosodic boundary at the end of the target phrase. The target word is thus always in prenuclear position.

Figure 1:

Examples of F0 trajectories over target phrases showing three accent patterns in stimuli: Primary (a), Early High (b), and Unaccented (c), as described in Table 1.

The speech materials were recorded in a sound-attenuated booth using a head-mounted microphone. To ensure that the three sound files for each sentence were identical except for the prosody of the target phrase, the three recorded versions of the target phrase were spliced onto the same sentence continuation. For instance, the final set of experimental sound files contained three files for the stimulus sentences (3a), where each file started with a different recording of the realistic story, one pronounced with an Early High pattern (the ré alìstic stóry), one with the Primary pattern (the rèa lís tic stóry) and a third one where the initial content word was unaccented (the realistic stóry). The remaining part of each file was the same single recording of the sentence continuation for all three files.

The final set contained 36 stimulus sound files (the 12 sentences in (3) each produced with the 3 prosodic patterns) and was divided into three lists of 12 sentences each, such that each subset contained all 12 sentences from (3) and comprised four sentences in each of the three prosodic patterns. Thus, each sentence was presented in only one prosodic pattern in each list.

2.2 Participants and experimental task

Thirty-three participants for this experiment were recruited from the student population of a large US public university. All participants were monolingual, native speakers of American English (23 female and 10 male). Participants were asked to imitate aural stimuli, with the instruction to repeat the sentence they heard “in the way the model speaker said it.” The experiment took place in a phonetics laboratory. Stimuli were presented through headphones and participants also wore a head-mounted microphone. Each stimulus was presented two times in succession, in auditory format only (no text), after which participants were asked to repeat it. Participants were randomly assigned to one of the three stimulus lists, as defined in the preceding Section 2.1. The stimulus list was presented twice, resulting in two iterations of the imitation per participant. The initial iteration was subject to analysis, except in cases where speech errors occurred, in which case the second one was analyzed instead. The total duration of the experimental task was approximately 30 min per participant.

2.3 F0 measurement and modeling

A total of 396 imitated utterances were recorded (33 participants *12 utterances). Five imitations were excluded due to lexical or syntactic errors. The remaining 391 imitated utterances were subject to acoustic analysis using Praat (Boersma and Weenink 2020). The target phrase within each imitated utterance was manually segmented into three intervals. Interval 1 spanned from the start of the subject NP to the consonantal onset of the primary stressed syllable of the target word, interval 2 spanned from the end of interval 1 to the end of the prenuclear target word, and interval 3 spanned from the end of interval 2 to the end of the head noun of the target phrase, as shown in (4) with numbers 1-3 marking the right edge of these intervals. Within each of the three intervals, we extracted the maximum and mean F0, manually checking for and avoiding consonantal perturbations:

The imitated utterances were classified as Early High, Unaccented or Primary based on the relative difference of F0 mean values^[6] in each interval (1, 2, and 3), as in Table 2.^[7] This classification was based only on the F0 pattern produced in the imitated utterance, regardless of the intonational pattern of the stimulus that was being imitated. The classification was done in three steps as follows. The imitation was classified as Early High if 1 > 2; otherwise, it was classified as Unaccented if 3 > 2; otherwise it was classified as Primary. By this method, an imitation was classified as Early High on the basis of the mean F0 in the first two intervals (1 and 2), with no restriction on the F0 mean in 3. Similarly, an imitation not identified as Early High in step 1 was classified as Unaccented based on the presence of an increase in mean F0 from the second to third interval, without further consideration of the first interval. The remaining contours were counted as Primary, which included imitations in which the F0 contour was rising over the sequence 1–2, in which case 2 ≥ 1, and those with a flat F0 over 1–2, in which case 1 = 2 (within 4 Hz, see Footnote 7).

Table 2:

Classification of imitated F0 patterns, based on relative F0 maxima or mean values.

Early High	Unaccented	Primary
1 > 2	3 > 2	2 ≥ 1

After classifying the imitated utterances, the imitations of each prosodic pattern in the stimulus set were analyzed to determine how many were correctly imitated, and how many were imitated with either of the other two patterns in the classification scheme in Table 2. For example, this analysis calculated how many imitations of Early High stimuli were produced with an Early High pattern, a Primary pattern, and an Unaccented pattern based on the relative F0 mean (or max) values.^[8]

The classification analysis just described has the obvious drawback that it assumes that the imitated productions are adequately classified as instantiating one of the three intended patterns in the auditory stimuli. It is possible, however, that the best classification of the results would group the imitated F0 contours using a different number of clusters. To explore the possibility that participants actually produced a smaller or greater number of distinct F0 patterns, and/or the possibility they produced distinctions that are not captured by our three predetermined classes, we performed two clustering analyses over the F0 trajectories in the imitated utterances, unlabeled for the accent pattern of the stimulus. The first clustering analysis was conducted over the series of three mean F0 measurements that were the basis for the classification analysis described above. These three F0 measures define a coarse-grained, simplified F0 contour. The second clustering analysis was conducted using the full F0 contours from each imitated production, time-normalized to 30 samples using ProsodyPro (Xu 2013), with 10 equidistant samples in each interval 1, 2, and 3 shown in (4). The total number of imitations entered into each of the analyses included 129 imitations of the Primary pattern, 131 imitations of the Early High pattern and 131 imitations of the Unaccented pattern (total = 391). All clustering analyses were conducted using the k-means clustering algorithm for longitudinal data with the kml package in R (Genolini et al. 2015).^[9] As the kml analysis is based on the Euclidean distance between the mean of hypothesized clusters, we z-normalized subjects’ F0 values using the scale function in R (R Core Team 2024) to reduce differences due to gender and individual variation in pitch range.

The clustering analyses report the optimal number of clusters for the imitated F0 trajectories (modeled as time-series data), and the mean values at each time point for each reported cluster. We examined the optimal clusters from each analysis in relation to the F0 contours of the three PPA patterns, comparing the number of clusters and the mean F0 contour shape of each cluster.

2.4 Predictions

To the extent that the presence of a PPA is potentially informative about information structure, pertaining either to the discourse referent of the prenuclear word or of a downstream word (as shown by Bishop [2017]), we predict that participants will reproduce a PPA when it occurs on the target prenuclear word in the stimulus. Specifically, we predict that imitations of Unaccented stimuli will have no PPA on the target word, while imitations of stimuli with Primary and Early High PPA patterns will realize a PPA on the target word. As for the distinction between the Primary and Early High accent patterns, there is a weaker prediction. Previous accounts of the Early High pattern (cited in Section 2.1) make no claim about a meaning distinction between the Early High and Primary accent patterns for 2–1 words, suggesting that this distinction may truly be “ornamental” intonational detail. In light of findings from prior imitation studies discussed in Section 1 showing that memory for noncontrastive acoustic detail of pitch accents is less reliable than memory for contrastive accent specification, we predict that the distinction between the Early High and Primary accent patterns will be less accurately reproduced in imitations than the distinction between either of these and the Unaccented pattern.

3.1 Classification analysis

As described above, F0 mean values were extracted from the three intervals 1–3 for each target phrase, as defined in (4).^[10] The imitations were classified into three accent patterns based on the relative F0 mean values in the three intervals. Figure 2 shows the number of imitated utterances classified in each of the three accent patterns. A striking finding is the prevalence of the Early High pattern in the imitated productions, which accounts for fully 61 % of the imitated productions.

Figure 2:

Classification of imitated productions into three accent patterns based on the relationship between F0 means in three measurement intervals as described in Table 2. (The term “imitated accent pattern,” as depicted on the x-axis of Figure 2, will be employed consistently throughout the paper to denote the participants’ imitations, which in this figure were categorized into one of three accent patterns in accordance with the criteria outlined in Table 2.)

Next, we look at differences in F0 means between intervals 1 and 2 for each utterance produced by our subjects, classified according to the criteria in Table 2 above, regardless of stimulus pattern. As can be seen in Figure 3, in tokens classified as presenting a Primary pattern, the difference in F0 between intervals 1–2 tends to be negative; that is, the mean F0 in interval 2 is higher, indicating the presence of a high-tone pitch accent in the second foot of the target word. In production tokens that we have classified as Early High, instead, the mean F0 in interval 1 is higher than in interval 2. Finally, in tokens classified as Unaccented, differences in the F0 mean between the two feet of the target word cluster around zero.

Figure 3:

Difference in mean F0 between intervals 1 and 2 for imitated productions classified according to the criteria in Table 2 (x-axis). The red asterisk marks the mean of each imitated productions classified by the criteria in Table 2.

A linear mixed-effects regression was run in R (R Core Team 2024) with the packages lme4 (Bates et al. 2015) modeling variation in the difference in mean F0 between intervals 1 and 2 (1–2), with Accent pattern classified as in Table 2 (3 levels: Primary, Early High, Unaccented), participant self-reported Gender (2 levels: female, male) and participant Group (3 levels: A, B, C) as fixed factors, and random intercepts for Participant and Utterance. Significant effects were evaluated based on p-values calculated using Satterthwaite’s method for degrees of freedom. Significant differences were found for Accent pattern when Early High was compared to the other two patterns (Primary: β = −27.81, SE = 4.33, t [268.55] = −6.42, p < 0.001; Unaccented: β = −21.63, SE = 7.76, t [360.61] = −2.79, p < 0.01). No significant effects of Gender (F [1, 79.07] = 0.01, n.s.) or Group (F [2, 79.79] = 0.05, n.s.) were found.

Turning now to the difference in mean F0 between intervals 2 and 3, this comparison clearly distinguishes Unaccented productions from the other two, as can be seen in Figure 4. In utterances where the target word is unaccented, the pitch accent in interval 3 is not downstepped, since there is no preceding pitch accent, and the F0 mean of interval 3 is substantially greater than that of interval 2 (yielding a negative difference for the subtraction 2–3). A linear mixed-effects regression on the difference in mean F0 between regions 2 and 3, with the same factor structure as above, returns a significant difference between Unaccented and each of the other two Accent patterns (Primary: β = 30.65, SE = 8.54, t [355.85] = 3.59, p < 0.001; Early High: β = 17.28, SE = 8.44, t [346.36] = 2.05, p < 0.05). There were no significant effects of Gender (F [1, 40.86] = 1.72, n.s.) and Group (F [2, 41.19] = 0.12, n.s.).

Figure 4:

Difference in mean F0 in intervals 2 and 3 (see Table 2) for imitated productions (x-axis). The red asterisk indicates the mean of each imitated productions classified by the criteria in Table 2.

The classification of participants’ productions in three groups using the criteria in Table 2 thus results in three statistically distinguishable groups of data, differentiated by the F0 pattern over the prenuclear and nuclear accents of the complex subject noun phrase. This is unsurprising, since the classification of imitations into the three accent pattern categories was based on observed differences in F0 across the three intervals in the target region.

Next, we compare the classification of each imitated production with the accent pattern of its corresponding stimulus. What we find is that Early High stimuli were imitated with high accuracy across all productions (85 % accuracy), with lower accuracy for Primary (60 %) and Unaccented (15 %) stimuli (Figure 5). Figure 5 also shows the breakdown of incorrect imitations into the three designated accent patterns. As can be seen, Early High was also the most frequently produced “error” for both Primary and Unaccented stimuli. These errors are the basis for the high proportion of Early High patterns in the imitated productions, as shown in Figure 2. In other words, our participants’ clear preference for an early prenuclear accent resulted in most Early High stimuli being accurately imitated, but also in most Unaccented stimuli being imitated with Early High and a sizable proportion of Primary stimuli also being produced with Early High.

Figure 5:

Imitations grouped by stimulus accent pattern (x-axis). Colored bars indicate the accent patterns of the imitated production according to the relative F0 mean values in intervals 1–3 as described in Table 2. (The term “stimulus accent pattern” will be consistently employed throughout the remainder of this paper to designate the accent patterns of the auditorily presented stimulus sentences that participants were instructed to imitate. Figure 5 plots imitated productions grouped by the accent pattern of the stimulus that was the intended target of imitation.)

The analysis of the results that we have offered in this section has the drawback that it focuses exclusively on the relative mean F0 across three intervals, necessarily classifying all production in one of three groups. In the next section, we present the results of an automatic procedure that takes into account additional aspects of the intonational contour and allows for classification based on more detailed F0 trajectories.

3.2 Cluster analysis

Cluster analysis offers a data-driven exploration of the distinctions in F0 trajectories that participants produced in their imitations, without invoking a priori assumptions on the number or type of distinctions. In the first cluster analysis that we conducted, the mean F0 values in the three intervals of each production were submitted as a 3-point trajectory to k-means clustering.

Figure 6 shows that the 3-point F0 trajectories of the imitations are optimally categorized into four clusters using this algorithm. Three of the clusters obtained from this analysis can be identified with the accent patterns present in the stimuli: Early High (cluster A in Figure 6), Primary (cluster C), and Unaccented (cluster D). The analysis also returns a cluster where the F0 mean decreases, suggesting a pattern of downstep across peaks in the first and second interval (cluster B), a pattern that has characteristics of both the Early High and Primary accent patterns. Among these four clusters, cluster A (Early High) is the most common, accounting for 30 % of the data, followed by cluster B (Early High + Primary), which accounts for an additional 29 % of the data. Cluster C (Primary) and cluster D (Unaccented) represent 23 % and 18 % of the data, respectively. Combined, clusters A and B, where the highest mean F0 patterns occurs in the first interval, account for almost 60 % of all patterns produced in our imitation experiment, which is consistent with the results above in Figure 2.

Figure 6:

Output of kml cluster analysis over trajectories of three F0 mean values from intervals 1–3 of the imitated target phrases. Colored lines show the mean trajectories for each of the four clusters in the optimal clustering solution. Black lines are the input trajectories.

Additional cluster analyses further explore the distinctions in the imitated F0 trajectories, now with each trajectory submitted as a time series of 30 F0 samples drawn from the time-normalized F0 trajectory of the target phrase. The purpose of these analyses was to examine variation in F0 contour shape beyond the coarse-grained analysis based on the mean F0 in three intervals, presented above. We report first on a cluster analysis performed over the F0 trajectories in the target region of all the imitated utterances, with no grouping imposed. The optimal solution has only two clusters, shown in Figure 7. The mean contour of cluster A has an early peak (between time points 5–10, in the first interval) followed by a fall in F0 to a lower peak in the third interval (around time point 25), corresponding to the head noun of the target phrase, while the mean contour of cluster B has a peak in the second interval (between time points 10–15) followed by a rise to a higher peak near the end of the phrase (around time point 25, in interval 3). The mean contour of cluster A, which includes 56 % of the imitated contours, resembles the Early High contour of the stimuli (Figure 1b), while that of cluster B, with 44 % of the imitated contours, resembles a combination of the Primary and Unaccented contours of the stimuli (Figure 1a and 1c). These proportions are similar to the proportions from the classification analysis presented above (Figure 2) using the three pre-determined accent patterns of the stimuli, which yielded 61 % of imitations classified as Early High and the remaining 39 % split between Primary and Unaccented.

Figure 7:

Output of kml cluster analysis over trajectories of 30 time-normalized F0 values from target phrases of all imitated productions. Colored lines show the mean trajectories for the two clusters in the optimal clustering solution. Black lines are the input trajectories.

Three further cluster analyses were performed using the same 30-point F0 trajectories on subsets of the imitated productions grouped by stimulus accent pattern. The purpose of these analyses was to determine if there are regular patterns of variation in the imitated F0 contours for each of the stimulus accent patterns. The first observation from these subset analyses is that, for each stimulus accent type, the optimal clustering solution had two clusters, shown side-by-side in each row of Figure 8, of which one, arbitrarily labeled “A,” has the expected F0 trajectory shape corresponding to the stimulus accent type. For imitations of the Primary accent pattern, those in cluster A have the largest F0 excursion and highest peak in the second interval and a second, downstepped peak in the third interval (e.g., “rea | listic | story ,” with F0 peaks in bolded intervals, and underline marking the highest such peak). For imitations of the Early High accent pattern, the imitations in cluster A have a peak in the first interval and a much weaker peak in the third (e.g., “ rea | listic | story ”). Finally, imitations of the Unaccented pattern in cluster A have only one peak, in the final interval (e.g., “rea | listic | story ”). The A cluster includes a slim majority of the imitations for each accent type: 56 % of Primary, 53 % of Early High, 53 % of Unaccented.

Figure 8:

Output of kml cluster analyses over trajectories of 30 time-normalized F0 values from target phrases of imitated productions subsetted by stimulus accent pattern: Primary (a), Early High (b), and Unaccented (c). The optimal clustering of F0 trajectories in each group has two clusters, displayed separately in side-by-side graphs in each row. The red line marks the mean trajectory of each cluster, and black lines are the input trajectories.

The imitations grouped in the “B” cluster for each accent type do not conform to the expected F0 trajectories. Instead, considering the mean trajectory of each B cluster, it appears that the imitations in the B clusters don’t uniquely resemble any one of the three stimulus accent patterns. All three B cluster patterns show a weak F0 peak in the third interval. In addition, F0 is slightly elevated across the first and second intervals in the Primary and Unaccented data, while the B cluster pattern for the Early High data show a noticeable late peak at the very end of the first interval. An example pitch track of the B cluster pattern for the Early High in imitated productions is shown in Figure 9. For all three B cluster mean trajectories, the mean F0 of the first interval is higher than that of the second interval, which is the likely source of the high number of imitations classified as Early High based on the criterion or relative mean F0 (shown in Figure 6). Qualitatively, the mean trajectories of the B clusters can be described with reference to the Early High pattern (as shown in cluster A of the Early High data). Imitations in the B clusters of the Primary and Unaccented data resemble the Early High pattern realized with a damped peak in the first interval. Imitations in the B cluster of the Early High data suggest something different—where the peak at the end of the first interval may be interpreted as the leading tone of a falling accent (H+!H*) on the primary stressed syllable. In other words, the imitations in cluster B of Early High resemble the Primary accent pattern in having a pitch accent on the primary stressed syllable (e.g., “rea | lis tic | story ,” with falling pitch over “lis”) followed by a downstepped !H* accent (“story”). Henceforth we draw on this qualitative interpretation of the cluster B pattern of the Early High data and refer to this as the “H+!H*” pattern.

Figure 9:

Example pitch track of cluster B for Early High in imitated productions.

The observation in Figure 8 of two F0 patterns in imitations of Early High stimuli was based on the clustering analysis of 30 time-normalized F0 trajectories from the target word. Above we suggested an interpretation of the two F0 patterns in terms of the location of the pitch accent on the secondary-stress syllable for Cluster A, and on the primary-stress syllable in Cluster B. In principle, this difference in pitch accent location could also manifest in durational cues of the target word in imitation. We thus carried out further analysis to examine duration (milliseconds) of two intervals 1–2 from the target phrases of imitated productions. Figure 10 shows subsets of imitated productions categorized by stimulus accent pattern. Two clusters of duration for each subset of imitated productions are grouped by the output of cluster analysis of F0 contours described in Figure 8 above.

Figure 10:

Duration of the first two intervals (1–2) from target phrases of imitated productions, grouped by stimulus accent pattern: Primary (a), Early High (b), and Unaccented (c). This is further categorized by the output of kml cluster analysis presented in Figure 8. The red asterisk marks the mean duration of each interval.

From the subsets of imitated productions grouped by Early High (Figure 10b), we see that the relationship between intervals 1 and 2 is clearly different between clusters A and B, which correspond to the contour with H* tone on the secondarily-stressed syllable and the contour with H+!H* on the primarily-stressed syllable in Figure 8, respectively. Whereas interval 1 was longer than interval 2 for A cluster, interval 2 was longer than interval 1 for B cluster. This visual impression was further confirmed by statistical analyses using mancova in R (R Core Team 2024). There was a significant difference between clusters A and B in the Early High pattern (holding constant participants and utterances) on the combined intervals 1, 2, and 3 (F [3, 86] = 15.95, p < 0.001). Pairwise comparisons between the clusters were significant for intervals 1 (p < 0.001) and 2 (p < 0.005), but not for 3 (p = 0.2), suggesting that the duration of two clusters differed in intervals 1 and 2, but not in 3. Overall, these results are in alignment with those from the cluster analysis of F0 contours that imitation of Early High had two distinct patterns, i.e., the contour with H* on the secondarily-stressed syllable and a contour with an early F0 peak, which we suggest may be analyzed as H+!H* on the primary-stressed syllable.

A detailed examination of the F0 trajectories in the B clusters reveals additional features. As described above, visual assessment of the mean F0 contour of the B clusters in the Primary and Unaccented subsets finds a first peak in the end of interval 1 (Figure 8). These patterns for Primary and Unaccented trajectories in the B cluster differ from the cluster B pattern of the Early High subset, where there is a more pronounced F0 peak (i.e., higher peak, aligned later) in the end of interval 1, which may be interpreted as H+!H* anchored on the primarily-stressed syllable (in interval 2). Our duration analysis in Figure 10 supports this interpretation. The cluster B imitations for all three accent types have longer duration for interval 2 than interval 1, but this tendency is greater for the Early High subset (Figure 10b) compared to the Primary (Figure 10a) and Unaccented (Figure 10c) subsets. This visual observation was confirmed by statistical analysis using mancova with the same structure described above. Pairwise comparisons found significant differences between clusters B in Early High and the other two accent patterns on interval 2 (Primary: p < 0.05, Unaccented: p = 0.05), but not on interval 1 (Primary: p = 0.61, Unaccented: p = 0.77), suggesting that Early High differed from the other two accent patterns in the duration of interval 2, but not interval 1, for imitations grouped in cluster B. Pairwise comparisons also revealed non-significant differences between the B clusters of Primary and Unaccented on intervals 1 (p = 0.85) and 2 (p = 0.87), showing that Primary and Unaccented have similar duration in intervals 1 and 2 in these data.

Together with the results from cluster analysis of F0 contours, the duration results suggest that the imitations grouped into B clusters have a similar prominence pattern, with the exception that for the B cluster imitations of Early High, the primary stressed syllable in interval 2 has greater pitch and durational prominence than the corresponding syllables in interval 2 for the Primary and Unaccented data. The overall similarity across all B cluster imitations suggests that Early High is persistently activated in all imitated productions. We suggest that the differences between the B cluster imitations of each accent pattern is that, while all three B cluster patterns show an influence of the Early High pattern (i.e., a prominence asymmetry between intervals, 1 > 2), that influence appears stronger in the B cluster imitations of Early High stimuli. In that panel (8b, right), we observe the expected Early High asymmetry in F0 prominence (1 > 2), though now reduced due to the elevated F0 in interval 2 (compare to panel 8b, left). The influence of the Early High pattern on “incorrect” imitations of Primary and Unaccented stimuli (panels 8a and 8c, right), show little if any asymmetry in the relative prominence of intervals 1 and 2, considering both duration and F0. In other words, the “incorrect” imitations in the B clusters appear similar in being somewhat weak versions of Early High, i.e., with a weaker 1 > 2 prominence asymmetry that is characteristic of Early High.

To summarize, the subset clustering analyses for imitations of each stimulus accent type explains several key findings. First, the numerous imitations classified as Early High based on the relative F0 values over the three analysis intervals (Figure 5) have a likely source in the “inaccurate” (B cluster) imitations of the three stimulus accent types. Second, the appearance of a fourth F0 trajectory in the clustering analysis over the aggregated three-interval mean F0 values (cluster B in Figure 6), with downstep over the first and second intervals and lower F0 in the third interval is compatible with imitations in cluster B of the Early High subset (Figure 8b). Third, the B clusters of the subset analyses show two variants of the Early High—one where the “early” pitch accent in the first interval has a damped peak, and another where the F0 peak in the first interval seems to mark the start of a falling pitch accent in the second interval. Viewed in this way, we suggest that these imitations may have undergone a phonological re-analysis, mapping the Early High pattern of the stimulus onto a Primary pattern, but using a falling pitch accent rather than the rising accent of the Primary and Early High stimuli. Notably, this falling accent pattern with a late peak in the first interval was not represented in the stimuli.