1.3.1 Previous phonetic descriptions of Drehu prosody

Recent phonetic investigations on word level prosody and tonal marking in Drehu (Torres, Fletcher & Wigglesworth 2018; Torres & Fletcher 2020) have not been able to find acoustic evidence for word initial primary stress (Lenormand 1954) or secondary stress on the third syllable, reported by Tryon (1968). Instead, these studies suggest that a prominent syllable is marking the right edge of the prosodic word. Torres & Fletcher (2020) found that noun phrases are mostly realised with a rising tonal pattern, demarcating the prosodic word with a low tone (L) to the left and a high tone (H) to the right edge. The evidence shows that the initial low tone is consistently aligned to the left edge of the word. More precisely, it aligns between a function word and a content word, thus suggesting it is a phrase initial tone. Additionally, it was found that a word final H tone is aligned to the last full syllable of the word. This study also reports that tonal targets have similar terminal F0 values at normal and fast speech rate, showing no evidence of steeper rises when the speech tempo was increased, contrary to predictions of a strict segmental anchoring hypothesis (Arvaniti, Ladd & Mennen 1998).

A preliminary study on words embedded in sentence frames examined words realised in non-utterance final position (Torres, Fletcher & Wigglesworth 2018). This study found that there is a preference to make the right edge prosodically prominent through means of a word final H tone and wider pitch excursions to the right edge. In contrast to expectations set by previous descriptions (Lenormand 1954; Tryon 1968), no evidence was found for an acoustically salient word initial syllable, but instead a small increase in duration was reported for word final vowels.

2.5.1 Analysis procedures

The recordings were manually transcribed and then force aligned in WebMAUS, using a grapheme to phoneme conversion, with a parameter model based on SAMPA (Kisler, Reichel & Schiel 2017). This alignment process provided a textgrid for each utterance in which all phonemes were segmented and marked with boundaries. Since automatic forced alignment does not detect phonological processes in Drehu, and is otherwise also prone to alignment errors, manual correction was required. All utterances were thereafter visually inspected in Praat 6.0.48 (Boersma & Weenink 2017) and the segmentation of phones in target tokens was corrected when necessary. During the correction of segmental alignment, special attention was paid to the setting of phone boundaries. A boundary was set between vowels and nasals, laterals or approximants at the point where sudden changes in both amplitude and formant structure occurred. In case the change in formant structure was gradual, the segment boundaries were marked at the midpoint of the transition from vowel to liquid or approximant. For obstruents, the start of closure was marked as the onset and the start of high amplitude periodicity was marked as the onset of the next vowel. Fricatives were marked at the onset and offset of high amplitude aperiodicity (Harrington 2010). Pauses that were inserted prior or after the target tokens were also identified and marked. Pauses were visually identified in the spectrogram based on the acoustic signal, when no speech was uttered. In ambiguous cases when a pause could be realised prior to a stop closure, it was decided that the stop closure would be marked at 50 ms before the release of the burst, unless otherwise indicated by the presence of formants or an F0 trace.

Contrary to earlier descriptions in the literature (Lenormand 1954; Moyse-Faurie 1983), and similar to Torres & Fletcher (2020), variability regarding a strict CV syllabification in word final syllables was found. Words where we would predict epenthesis to occur were often realised without the epenthetic vowel ending on a coda consonant instead. This caused some variability in the number of syllables contained in token words. For example, a disyllabic word such as trenge [t͡ʃe.ŋe] ‘basket’, could instead be realised as the monosyllabic word [t͡ʃeŋ]. There were five words that showed this variability and a slight preference for epenthesis realisation was noted since 54.7% of the tokens were realised with a final vowel.

For query purposes within emu-R (Winkelmann et al. 2017), the following parameters were marked: position in carrier phrase (Initial, Medial, Final), position of the syllable, and pauses inserted prior or after the target token. Additionally, an AM-style annotation (Beckman & Hirschberg 1994; Pitrelli, Beckman & Hirschberg 1994; Welby 2006; Ladd 2008) was used to mark tonal targets in the items of interest. The F0 contour was visually inspected in Praat and the tones were marked according to fundamental frequency minima and maxima points that were found. The location of L was defined as the F0 minimum, and mostly occurred around the first syllable of the test word. High tones were also visually identified and marked with H. When more than one rising tone was observed, the tones in the initial rise were marked with an additional letter (L_iH_i) and every subsequent tone in a token was also marked with an additional number (L₁H₁). In case there was a plateau, valley or several points of equal F0 value, the tonal target was marked at the onset of the tone. Attention was given to contexts when the word initial consonant was an obstruent or a fricative, and tonal targets were marked to avoid the marking of micro-prosodic perturbations as tones. In some instances, especially for words in sentence final position, it was difficult to determine a tonal target due to strong glottalisation or a very breathy voice quality. In these cases, it was preferred not to assign any value to the final tone target.

Although there was variation in the intonational configuration of patterns, four tonal movements that could be associated with phrasal prosodic boundaries, were identified and analysed. Three types of rises and one falling tone were observed in the data which, for the sake of clarity, are referred to as: the left-boundary rise (LiHi), the across-token rise (LH), the right-boundary rise (L₁H₁), and falling tone (HL). These movements were sometimes combined on a single word, see Figure 2. A small proportion of target items produced with disfluencies or noise (e.g. coughing) were discarded, leaving 91% of the data in our corpus. A total of 751 utterances, that were mostly evenly distributed amongst speakers, were collected and used for analysis.

Figure 2

Schematic representation of seven tonal patterns identified in the Drehu corpus.

Figure 3 shows the notation used for the data and how target tokens, position, syllables, phones, and tones were labelled. A hierarchical database was constructed using the EMU Speech Database Management System (Winkelmann, Harrington & Jänsch 2017). It included five tiers for the tones, phonemic segments, syllables, words, and target token position. The acoustic and durational characteristics produced in the target words were extracted and analysed using the emuR package in R (R Core Team 2017; Winkelmann et al. 2017). First, the tonal patterns were noted, then duration values for syllables, and F0 values of low and high tones were extracted. For further analysis, the values obtained in Hz for tonal targets were converted into semitones with a benchmark of 100 (Nolan 2003). Since we were interested in the magnitude of rises, we measured pitch span expansion as the difference between two consecutive tonal targets (T1-T2) in semitones. Following Campbell (1999), in order to examine variation related to a prosodic order, durational values obtained for syllables were normalised to z-scores per individual speaker. Expressing duration on a scale from long to short, with the average being the centre point, the z-score transform is known for factoring out predictable contextual differences due to inherent length of segments. Thus, predictable effects related to longer duration of segments can be factored out by calculating the averages separately for each syllable. This makes it possible to consider lengthening as a result of prosodic effects.

Figure 3

Speech waveform, spectrogram, and F0 trace of target token ahnahna ‘gift’ highlighted in a black box, from top to bottom, inserted in sentence Initial, Medial (with following pause), and Final (after a pause) position. All examples were produced by the same female speaker.

2.7.1 Tonal patterns

First, collected data, tonal patterns and observations regarding the insertion of pauses will be presented. Sentences that contained misspellings or hesitations in the critical target word were discarded from further inspection. As mentioned above, due to strong glottalisation, devoicing or a breathy voice quality, not all tones could be identified, meaning that some words could not be tonally marked, especially in sentence final position. In the corpus, tonal patterns were identified for 606 target words, in three positions: sentence Initial (227), Medial (205), and Final (174). Figure 2 shows a schematic representation of the seven different tonal patterns found. In comparison to Torres, Fletcher & Wigglesworth (2018), one additional pattern was identified (L_iH_iH₁), although this represents a low frequency pattern (1.8%).

Figure 4 shows the occurrences of tonal patterns identified in our corpus according to position in the carrier phrase. Figure 3 (top panel) shows the pitch trace of a LH rising pattern, the most frequent in the corpus, found in 46.8% of the entire data set. This rising contour also represents the across-token rise (LH). The LL₁H₁ pattern represents the second most frequent rising pattern, occurring in 12% of the cases. Note that the F0 trajectory of LH in Figure 3 (top panel) differs from that in LL₁H₁ in Figure 3 (mid panel). While LH displays a gradually ascending F0 curve, the pattern of LL₁H₁ is characterised by a low plateau, sometimes with a slight dip, which then rises to a peak at the right edge of the prosodic word (and of the Intonational Phrase in this example). The LL₁H₁ tonal pattern illustrated in Figure 3 (mid panel) contains the right-boundary rise (L₁H₁). Three other patterns that ended on a high peak at the right edge of token words were identified: HLH₁ (7.8%), L_iH_iL₁H₁ (4.6%), and L_iH_iH₁ (1.8%). Additionally, two tonal patterns that ended on a low target were also identified, HL (26%) and L_iH_iL₁ (1%). These two patterns carry a falling tone (HL) to the right edge. As can be expected, most of the patterns ending on a low tone occurred when the token word was in sentence Final position. However, the pattern HL was also identified in sentence Initial and Medial positions a total of 12 times, and in 5 cases it was followed by a pause.

Figure 4

Bar plots summarising the tonal patterns identified in the corpus according to position in the carrier phrase. There were three positions Initial, Medial, and Final. The plot includes pre- and post-pausal tokens.

From the 174 patterns identified in sentence Final position 84% were realised with a HL pattern. However, note that the LH rising pattern was also identified in sentence Final position, occurring in 13.7% of the data. Additionally, the L_iH_iL₁ and LL₁H₁ were identified, each in 1.15% of tokens in sentence Final position.

It was noted that participants sometimes inserted pauses, as illustrated in Figure 3 (mid panel), where a pause was inserted after the target item, and Figure 3 (bottom panel) with a pause occurring prior to the target item. The frequency of the insertion of pauses was quantified and it was found that in 22.5% of the data a pause was inserted prior or after the target word. Sentence Initial target words were realised in pre-pausal position in 6.8% of the cases. Sentence Final target items were in post-pausal position in 15.3%. Interestingly, sentence Medial target words were accompanied by a pause in 44.5% of the cases, in 92.4% the target word was realised in pre-pausal position and only in 7.6% in post-pausal position.

2.7.2 Tonal movements

In addition to examining the distribution of tonal patterns in the corpus, in this experiment the magnitude of tonal movements was examined. This was done in order to verify whether greater expansion could be signalling a higher prosodic level, such as the IP. For this purpose, the magnitude of pitch span in rising and falling tones was calculated as T1-T2. With this procedure, rising tones receive a pitch span measurement with a positive value, while falling tones a negative one. Table 1 shows the mean values for pitch span expansion between tonal targets in three positions. These values indicate that in absence of a pause, tonal movements in sentence Final position appear to be greater in magnitude than in sentence Initial and Medial positions.

As mentioned earlier, target items in sentence Final position were mostly realised with a HL pattern, although also a LH pattern was reported, and only in less than 3% of the data L_iH_iL₁ and LL₁H₁ were observed in this position, see right panel in Figure 4.

In the following, the aim is to examine the use of F0 in demarcating prosodic boundaries. Therefore, we re-examine tonal patterns to see which ones a) demarcate the right edge of accentual phrases, and b) higher level tonally-marked units like intonation phrases. Additionally, we examine the F0 realisations of these tonal movements in more detail including pitch span (i.e. the distance between T1 and T2).

For instance, it could be that in the HL and L_iH_iL₁ patterns the word final L is always associated with a major prosodic boundary. Figure 5 shows the magnitude of pitch span (T1-T2) of four tonal movements observed in three positions (see also Table 1). Note that the falling tone (HL), was a low frequency pattern in Initial (4 occurrences by 3 speakers) and Medial (8 occurrences by 3 speakers) positions, while it was more frequent in Final position (146 occurrences). Therefore, mean values should be treated with caution. As shown in Figure 5, the L tone in HL in sentence Final position, is lower than in other positions. This L tone is thus associated with final lowering, an additional lowering relative to other Ls earlier in the phrase. Moreover, the deeper falls (around – 6 st) could be additionally affected by declination. Taking into consideration measurements in Table 1, this shows that both tonal movements, the across-token (LH, 8.7 st) and falling tone (HL, –9.4 st) in sentence Final and pre-pausal position are realised at the extremes of the tonal range of speakers.

Figure 5

Box plots show magnitude of pitch span in semitones for tonal movements in sentence Initial and Medial position and (right-edge tonal movements) in Final position. Pre-pausal and post-pausal items excluded.

Table 1 and Figure 5 show that sentence position and movement type are important factors in determining pitch movement magnitude so a linear mixed effects model was employed to investigate whether the position in the sentence had an effect on the magnitude of the tonal movement. The model had Position (Initial, Medial, Final) and Tonal movement (left-boundary, across-token, right-boundary, falling tone) as fixed effects, together with participant and token as random intercepts, no random slopes were included. From the initial 606 annotated intonation patterns 134 co-occurred with a pause, and tokens in pre-pausal and post-pausal positions are excluded from this analysis. In 70 cases, due to pitch track errors, we could not obtain a value for one of the two tones required to measure pitch span, leaving thus 421 data points for further analysis. The reader is reminded that in e.g., L_iH_iL₁H₁, two rising tones are present. There were 19 such instances for which two tonal movements (left-boundary and right-boundary) were included. In the case of tri-tonal patterns, of the type L_iH_iL₁, where one tone could be associated with a rising and a falling tonal movement at the same time, both tonal movements were included. Since we did not want to discard tonal movements a priori, it was decided that both the left-boundary and the falling tone would be considered. Additionally, for the pattern L_iH_iH₁, the left-boundary and for LL₁H₁ the right-boundary tonal movements were taken into account, as we did not expect any strong differences between two consecutive low and high tones, and these were low frequency patterns.

Results of the linear mixed effects model are in Appendix B.1. After performing a backward model selection, Position and Tonal movement were retained as fixed factors, together with random intercepts for participant. This model was used for a post hoc analysis which examined an interaction between the two retained fixed factors.

Selected results obtained from the Tukey corrected post hoc tests are summarised in Table 2. They show that there is an interaction between position in the sentence and tonal movement but crucially that for the across-token rise there is a significant difference in sentence Final position when compared to Initial and Medial positions. This shows that the magnitude of pitch span is substantially greater when in sentence Final position. As expected, no significant effect was found when comparing the across-token rise in sentence Initial and Medial positions. However, no significant difference is found when the Final falling tone is compared to Initial and Medial positions. As stated before, the HL pattern was a low frequency pattern in Initial and Medial positions (7 occurrences in absence of a pause). Considering that the model included 11 data points for falling tones, we take this result with caution as additional data points would be necessary to contribute to a more definitive result.

Arguably, for sentence Final tokens the position in which they were realised is already indicative of their right boundary being at a higher prosodic level, such as the Intonation Phrase. Given that tones in sentence Final position are associated with phrasal boundary marking, demarcating a higher prosodic level, it should be no surprise when larger differences are measured in this position. Considering the negative values for the falling tone in sentence Initial and Medial position, we take HL and L_iH_iL₁ to be patterns whose right boundary tone (L) can be associated with a boundary of a higher prosodic level, namely the IP. When occurring in utterance final position, the notation L% and H% should be used to indicate they are IP-boundary tones.

2.7.3 Rising tones and pauses

Pre-pausal and post-pausal rising tones in sentence Initial and Medial positions were of interest, since the insertion of a pause could be signalling a higher prosodic level, namely the IP. Therefore, the pitch span between the low and the high tones associated with these two contexts was compared. As mentioned in section 2.5.1, three types of rises within the word were established, the left-boundary rise (LiHi), across-token rise (LH), and right-boundary rise (L₁H₁). Box plots in Figure 6 show the differences in pitch span for the three rises when realised without a pause and in pre-pausal position.

Figure 6

Box plots show magnitude of pitch span in semitones for three types of rises found in Initial and Medial positions. Tokens realised with a following pause are in dark grey.

To investigate factors influencing the magnitude of the excursion a linear mixed effects model was employed. The model included Position (Initial, Medial, Final), Rise type (left-boundary, across-token, right-boundary), Pre-Pause (yes, no), and Post-Pause (yes, no) as fixed factors as well as participant, and token as random factors, no random slopes were included.² The model included 401 observations, from which 96 were produced in pre-pausal and 5 in post-pausal position. Additionally, for 22 words two rises were included, and the remainder were taken from 357 patterns with only one rise.

The results of the linear mixed effects model are in Appendix B.2. They show that rises of post-pausal words were not significantly affected by the pause. Thus this factor was excluded from the subsequent model. After performing a backward model selection, the factors Position, Rise type, and Pre-Pause were retained together with random intercepts for participant and token.

Selected results for the Tukey corrected post hoc test are summarised in Table 3. Results are averaged over the levels of position. The analysis reveals that there was a significant effect of pause on rise type, and that the pitch span expansion was greater when tokens were followed by a pause. This effect was significant for the across rise and for the left-boundary rise which decreased when in pre-pausal position, see also Table 1. Additionally, compared to the left rise (L_iH_i), the magnitude of the excursion of the right boundary rise (L₁H₁) is significantly greater. Moreover, there is a significant difference between the left rise (L_iH_i) and the across-token rise (LH), with the latter also being significantly greater. Interestingly, there is no significant difference in the size of excursion for the across-token and right-boundary comparison. These results suggest that the across-token and the right-boundary rises show similar strength and contribute to demarcating the right boundary of the AP or the IP when in pre-pausal position.

2.7.4 Word duration

An additional aim of this experiment was to determine whether the manipulation of position in the sentence had an effect on duration patterns of the word and the constituent syllables. Table 4 provides a summary of word length in number of syllables per word, and token numbers in three different sentence positions. Figure 7 shows the duration of words with one, two, three or four syllables in three different positions in ms. As expected the acoustic duration of words increases incrementally depending on word length in syllables.

Figure 7

Box plots show duration values in ms of words with one to four syllables in three different positions in the sentence: Initial, Medial, and Final. Words in pre- and post-pausal position were excluded.

Interestingly, words with identical syllable numbers are longer the further to the right (or later in the sentence) they are produced, especially in sentence Final position. A linear mixed effects model was employed to investigate the effect of position on token duration. From the 751 collected utterances, 166 target words co-occurred with a pause and were excluded from this analysis, leaving 585 observations. The model included Position (Initial, Medial, Final) and Syllables (monosyllable, disyllable, trisyllable, quadrisyllable) as fixed factors together with the random factors participant and token, no random slopes were included. Results of the linear mixed effects model are summarised in the Appendix B.2.

After performing the backward model selection, the model retained the fixed factors Position and Syllables with random intercepts for participant and token. Selected results of the post hoc analysis are summarised in Table 5, they indicate that the position in which, e.g., disyllabic words were produced, had an effect on their duration, and that words produced in sentence Medial and Final positions are significantly longer than words in sentence Initial position. Not surprisingly, as word length increases in terms of number of syllables, acoustic duration also increases. However, there is no significant difference when monosyllables and disyllables are compared in sentence Initial, Medial, and Final positions.

2.7.5 Syllable duration in the word

Apart from identifying whether the manipulation of position of the word had an effect on its duration, the aim was to examine whether there is evidence for an acoustically more salient syllable in the word. For this purpose, syllabic duration in polysyllabic words was examined. A linear mixed effects model was employed to investigate syllable duration of polysyllabic words, including observations from 1267 syllables. To allow for better comparison and avoid confounding effects, this analysis was restricted to syllables with a cv-structure, taken from 524 words (191 disyllables, 220 trisyllables, 113 quadrisyllables). Tokens in pre- or post-pausal position were excluded. Normalised duration values were fitted to a model that included the Position in sentence (Initial, Medial, Final) Position of the syllable (first, second, third, final) and Syllable number per word (disyllable, trisyllable, quadrisyllable) as fixed factors, together with participant and token as random factors, no random slopes were included. Results of the linear mixed effects model are in Apendix B.3.

Table 6 shows the mean duration of cv-syllables in words with varying syllable number produced in the three positions included in the experiment. Note that for disyllabic and trisyllabic words mean duration for syllables with an epenthetic vowel is shorter than for other final syllables. Although in this data set only 23 words were realised with epenthesis it was decided to exclude these items from the following statistical investigation. Note that mean duration values in Table 6 indicate that syllables in polysyllabic words are shortened.

After performing backward model selection, the model retained the fixed factors Position in sentence, Position of the syllable, Syllable number per word and token, as random factor. Of particular interest was the evaluation of a possible interaction between the number of syllables in the word and the position of the syllable in the word. Additionally, an interaction between the position in the sentence and the position of the syllable in the word was examined. Selected results of the Tukey corrected post hoc analyses are summarised in Table 7.

The analysis shows that the duration of a syllable is influenced by its position in the word and the position of the word in the sentence. There is a statistically significant interaction for the position of the syllable in the word and the number of syllables per word in polysyllabic words with three and four syllables. However, this result was not the same in words with only two syllables. Similarly, there is a significant interaction for the position of the word in the sentence and the position of the syllable in the word, showing that word final syllables are significantly longer in sentence Final position when compared to Initial and Medial positions. In contrast, no significant difference is found for word final syllables in sentence Initial and Medial positions.

2.7.6 Additional effects on syllable duration

Similar to the analysis of the effect of pauses on rises, it is of interest to determine how syllable duration was affected by the insertion of a pause. Figure 8 visualises cv-syllable duration of disyllabic, trisyllabic, and quadrisyllabic words produced without (no) and with a pause (yes). The box plots indicate that when produced with a pause, the target word was longer but specially that the final syllable appears to be lengthened.

Figure 8

Box plot of normalised syllable duration in words containing two, three, and four syllables. Produced without (no) and with a pause (yes). Items from three positions in the sentence (Initial, Medial, Final) are included.

To further investigate syllable duration and a possible interaction with the insertion of a pause a linear mixed effects model was employed. The model included measurements of 1665 syllables, taken from 678 words. In this sample 38 words were in post-pausal, 116 in pre-pausal position, and 524 did not co-occur with a pause. Based on the previous sets of analyses we tested the fixed factors Position of syllable (first, second, third, final), Position in sentence (Initial, Medial, Final), Pre-Pause (yes, no), Post-Pause (yes, no) and token as random intercept, no random slopes were included. The results for the model are in Appendix B.5. After performing backward model selection, the model retained the factors Position of syllable, Position in sentence, Pre-Pause, and token as random intercept. Selected results for the Tukey corrected factor comparisons between Pre-Pause and Position in the sentence as well as Pre-Pause and Position of syllable in word are summarised in Table 8.

The analysis reveals that the insertion of a pause affects tokens in the same position and that when a pause is inserted, there is lengthening. Note that the previously established difference between duration values of words realised in sentence Initial and Medial positions, is again confirmed for the normalised duration of syllables. The results further indicate that the final syllable was significantly lengthened, when a pause followed, in comparison to when no pause was inserted. Additional evidence comes from the comparison between the third and the final syllable, whereby the final syllable is significantly lengthened when produced prior to a pause. There is no effect of pause on non-final syllable duration, i.e. non-final syllables in pre-pausal tokens do not lengthen. Taking into account the previous intonational analyses, it appears to be the case that sentence Final and pre-pausal word-final syllables that carry an IP-final boundary tone (L% and H%) are also significantly longer, thus providing non-tonal evidence of a higher level prosodic boundary.