Five local participants in each dialect were recruited as the representative speakers of Changsha dialect (M_age = 62.00 yrs., SD = 6.36 yrs.; 2 females, 3 males), Chengdu dialect (M_age = 58.60 yrs., SD = 7.33 yrs.; 1 female, 4 males), Fuzhou dialect (M_age = 62.40 yrs., SD = 6.50 yrs.; 2 females, 3 males), and Xiamen dialect (M_age = 61.40 yrs., SD = 8.73 yrs.; 2 females, 3 males). In total, 20 participants took part in this experiment. Consistent with the traditional manner of in-depth field investigation, we only chose the participants aged 50 and older as representatives of each dialect. The reason is that the phonology of nowadays young people is often greatly influenced by Beijing Mandarin (Yao and Chang, 2016).

All participants were born and raised in the downtown of the local cities without the experience of traveling outside for over 6 months. According to the self-report, they only acquired their native dialect without the experience of other Chinese dialects or foreign languages. This effectively avoided potentially prosodic influences from other dialects/languages (Archibald, 2009). Besides, they did not self-report any speech disorder or hearing impairment. Before the elicitation task via text prompts, we confirmed that all participants could read the text of Chinese characters normally. After the experiment, each participant was paid the equivalent of 15 USD in local currency for their travel and time.

The experimental stimuli, listed in Supplementary Table 1 (due to its overlength), were disyllabic prosodic words partially chosen from the word list (Guo, 2020) designed to investigate tone realizations from a cross-dialectal perspective. The lexical items were compound words including nouns, verbs, and adjectives, which were frequently spoken words [such as “工人” (worker)] across all the four dialects. Besides, we also selected some colloquial words in each dialect from previous studies as supplements, such as /ts^hu53 ting53/ (rooftop) used in Xiamen dialect only (Chen, 1987). Therefore, in the current study, we used a mixed word list with both dialect-universal and dialect-specific words.

To make sure that the tone sandhi could be comprehensively analyzed, the disyllabic lexical items contained all the possible tonal combinations in each dialect, and five lexical items were chosen under each tonal combination as tokens (Supplementary Table 1). Since the number of lexical tones is different in the four dialects, the number of tonal combinations is also different. For instance, in the four-toned Chengdu dialect, the number of tonal combinations is 4^*4 = 16, with a total of 16^*5 = 80 lexical items. As for Changsha dialect, the total experimental items were 6^*6^*5 = 180 words. In Fuzhou and Xiamen dialects, however, the Tones 6, 7 belong to the checked tones (or “Ru Tones”). First, they are naturally shorter than other lexical tones (Tones 1–5) in terms of duration. Moreover, the checked tones are also different from other lexical tones with unique coda such as glottal stop /-/ or consonant stops (/-p/, /-t/, and /-k/). When they precede other lexical tones, its coda may drop which causes the compensatory lengthening of the duration (Chen and Norman, 1965). To control variables and make duration analysis more comparable, the checked tones (Tone 6 and Tone 7) in Fuzhou and Xiamen dialects were not included in the current study. Thus, there were 5^*5^*5 = 125 lexical items for both Fuzhou and Xiamen dialects.

The recordings were conducted in quiet rooms located in Changsha, Chengdu, Fuzhou, and Xiamen city, respectively. To record high-quality audio samples, we used the cardioid microphone (AKG-C554L) connected to a USB audio interface (iCON4 nano VST). The recording for each dialect was conducted with a relatively low-level environmental noise (under 30 dB SPL). Before the formal recording, all the lexical items were shown to participants to familiarize them with the recording materials. Besides, a pilot investigation was carried out to confirm that five disyllabic prosodic words under each tonal combination showed the same tonal representation.

To control the pronunciation variables, a carrier sentence “这是__ (This is __)” was used before the target word. Participants were asked to produce the pre-target carrier sentence and the target word in a natural manner. All the target words were presented in a random order among participants. Specifically, participants could see the Chinese characters and their related lexical meanings of the target word on a laptop screen. Then, they spontaneously produced both the carrier and target word three times based on these prompts. All the participants of each dialect correctly uttered the target words, and the signals were saved in a WAV format with a sampling rate of 44.1 kHz.

After the collection of raw data, all the sound files were processed by the Praat (v.6.0.26) (Boersma and Weenink, 2021) on a PC laptop. Due to the occasionally poor sound quality (i.e., creaky voice) spoken by some participants, the pitch tracking might drop out throughout the syllable. To obtain more reliable results, only the best sound file with continuous pitch contours (exhibited in Praat) for each lexical item was included in the statistical analysis. In some cases, all the three recorded samples of the target word showed pitch-tracking failure, then we manually fixed the pitch tracking by the pitch tier function in Praat (Styler, 2011). In sum, there were 900 analyzed tokens in Changsha dialect (5 speakers × 180 words), 400 tokens in Chengdu dialect (5 speakers × 80 words), and 625 tokens in both Fuzhou dialect (5 speakers × 125 words) and Xiamen dialect (5 speakers × 125 words), with 2,550 analyzed tokens in total.

The duration of each syllable was measured from the finals such as vowel (V), nasal rhyme (VN), and rhyme with glide (GV), since these different finals in Mandarin would not cause a significant duration difference (Wu and Kenstowicz, 2015). These finals were manually identified based on the spectrogram information in Praat, i.e., the onset and offset of the second formant (F2) within finals (Turk et al., 2006). Then, the raw duration was normalized for each participant with the z-score method according to Rose (1987). Since the duration is often skewed in distribution, the normalized duration was log-transformed as the dependent variable when entering the statistical models. In addition, to compare duration contrast among four dialects, we also calculated the σ(1) to σ(2) mean ratio of the absolute duration.

The fundamental frequency (F0) was extracted from each manually labeled syllable, in which 11 equal-distance points for the pitch trajectory were outputted. These F0 points were further checked and manually corrected for any “pitch-halving” or “pitch-doubling” errors which are detected when the determined F0 value is 20% higher or lower than the reference F0 value (Sun, 2002). Then, the raw F0 values (in Hz) by each participant were transformed into the logarithmic z-score values to eliminate individual differences in pitch range (Zhu, 2010).

All acoustic data were analyzed using R (R Core Team, 2020). To compare the duration between two consecutive syllables, a one-way ANOVA was conducted for each dialect. Moreover, linear mixed-effect models were conducted with the lme4 package (Bates et al., 2015). The p-values of fixed factors and their interaction were obtained with a type-II ANOVA using Wald chi-square tests via the car package (Fox and Weisberg, 2019). Furthermore, second-order orthogonal polynomials were built to compare three parameters of pitch contours (Mirman, 2014): the intercept term (i.e., overall pitch height), the first-order linear term (i.e., pitch slope), and the second-order quadratic term (i.e., pitch curvature).

The random slopes and intercept were incorporated in all models to make it generalizable across data maximally (Barr et al., 2013). Then, model comparisons were conducted to find out the best-fit model based on the Akaike information criterion (AIC) using the MuMIn package (Bartoń, 2022). When a significant main effect of a multilevel factor or a significant interaction effect was found, post-hoc pairwise comparisons were performed by using the lsmeans package (Lenth, 2016) with Tukey adjustment.

Generally speaking, the duration could be used as a phonetic cue measuring the syllable weight, since a longer duration might indicate a heavier syllable weight (Maddieson, 1993; Hubbard, 1994). Especially in the weight-sensitive languages, the heavy syllables carrying two morae exhibit a longer duration than the monomoraic light syllables, and also are easier to attract the metrical stress (Selkirk, 1982; McCarthy and Prince, 1994). The relation between the syllable weight and the word stress could be summarized as the “weight to stress principle” (Prince, 1990). The present study revealed the diversity of duration patterns between two types of metrical structure as shown in four Chinese dialects, and the results were discussed as follows.

According to our prediction (H1), the left-dominant structure might exhibit a long-short pattern in duration. Yet, our finding does not support the H1, since although the duration between two syllables was significantly different in Changsha dialect, that was evenly arranged in Chengdu dialect. Specifically, in Changsha dialect, the duration in σ(2) was significantly shorter than that in σ(1) regardless of tonal categories. This duration pattern was supported by the impressionistic description according to Zhong (2003). The mean duration ratio in Changsha dialect was about 1.53: 1, approaching the ratio (1.7: 1) of the neutral-tone words in Mandarin (Chen and Xu, 2006). It is worth noting that lexical items in our study were daily spoken words that did not belong to the category of the neutral-tone words used in a small number of Mandarin words (about 7% according to Li, 1981). However, the disyllabic words in Chengdu dialect exhibited comparable duration between two syllables (with a ratio of 1.03: 1), which pointed to the absence of a phonetic reduction in σ(2). Thus, the syllable weight indicated by duration was quite evenly matched in disyllabic words of Chengdu dialect. Given the left-dominant structure in Chengdu dialect according to Qin (2012), the duration may not act as the phonetic correlate of metrical prominence robustly. Potentially, the difference in duration pattern in these two dialects could be attributed to the dialect differences, since Changsha and Chengdu dialects belong to the Xiang dialect group and the Mandarin supergroup, respectively (see Figure 1A). It was reported that disyllabic words in the Yiyang dialect (Xiang dialect group) also showed a long-short duration pattern (Xia, 2018). More cross-dialectal research is needed to demonstrate the diverse duration patterns of the left-dominant metrical structure.

On the contrary, a short-long duration pattern was consistently found under the right-dominant structure of both Fuzhou and Xiamen dialects. Such duration pattern might serve as the robust phonetic evidence for the right-dominant structure, indicating that Fuzhou and Xiamen dialects were weight-sensitive dialects. This finding is consistent with our prediction (H1) for the right-dominant structure, and is also in line with the “Iambic/Trochaic Law” (Hayes, 1995), suggesting that a short-long duration pattern commonly exists in the iamb. Although Fuzhou and Xiamen dialects shared a similar duration pattern, the duration ratio difference between two syllables was not exactly comparable in these two dialects. To be more specific, the duration contrast between σ(1) and σ(2) in Fuzhou dialect (a ratio of 0.57: 1) was generally greater than that in Xiamen dialect (a ratio of 0.83: 1). When focusing on the duration contrast under different tonal categories, as shown in Table 2, we could also observe a greater duration difference between σ(1) and σ(2) in Fuzhou dialect. A deeper understanding of this phonetic difference in the right-dominant structure requires further investigation.

In general, duration patterns in three out of four dialects corresponded with the underlying metrical structure. Thus, the metrical prominence in disyllabic words of Chinese dialects might not always be manifested as a longer duration. Such finding only supported our prediction (H1) regarding the right-dominant structure. In addition, consistent with the “Iambic/Trochaic Law” (Hayes, 1995), the duration pattern could reliably reflect the “relative relation” (i.e., light or heavy) of the syllable weight in the right-dominant structure of Chinese dialects. Future research needs to be carried out to investigate whether there is a final-lengthening effect on duration realization by putting the target words at different prosodic positions of the carrier sentence.

It was proposed that pitch realizations between metrically weak and strong positions in Chinese have opposite tendencies. One is that, in the metrically weak unit such as σ(w), the lexical tone tends to lose its underlying contour (Chen and Xu, 2006), or further undergoes tonal alternation (Yue-Hashimoto, 1987; Chen, 2000); The other is that the syllable with more prosodic strength [σ(s)] would fully exhibit the underlying tonal representation and keep its original tone features (Kochanski et al., 2003; Zeng and Niu, 2006). The current study investigated the pitch realizations of four Chinese dialects by using the second-order orthogonal polynomials models to compare the pitch intercept, slope, and curvature.

Table 7 shows tonal alternations in the metrically weak position across the four dialects. Specifically, there were generally three trends of tonal changes according to underlying pitch contours: (1) The underlying rising tone was generally realized as a level tone on the surface (i.e., [44]/[33]/[22]). Moreover, the pitch height of the surface level tone was below the peak pitch of the underlying form, except for Fuzhou dialect (no underlying rising tones in its tone inventory); (2) For underlying level tones, they were prone to change into falling tones with similar pitch height. Note that there are no underlying level tones in the tone inventories of Chengdu and Changsha dialects; (3) The underlying falling tones tended to be realized as level tones with similar pitch heights (i.e., [33]/[44]), or sometimes be realized with original falling contour but different pitch heights. In summary, a notably consistent tonal pattern was that underlying contour tones (falling or rising) were generally realized as level tones on the surface. In other words, the level tone was a more common surface representation in the metrically weak position.

Moreover, it should be noted that these surface tones were sometimes context-dependent in the four dialects. In other words, the tone sandhi in a metrically weak position might be conditioned by the preceding/following tonal context (of metrically strong position). For instance, the pitch realizations of T2-T2 in Chengdu dialect and most tonal alternations in Fuzhou dialect were contextually conditioned. Specifically, in Chengdu dialect, the surface representation of T2 in σ(2) was realized as [33] only when preceded by another T2 in sequence (see Figure 5). Likewise, the surface pitch form of T2 in σ(1) was decided by the following tonal context in Fuzhou dialect (see Figure 6). These tonal changes are overall comparable to the neutral tones in Mandarin, where tone deletion occurs, leaving a vacant tone-bearing unit in the target position which awaits a proper pitch realization from the preceding tonal context (Wang, 2008). Likewise, the metrical weak unit might contain a single mora. Additionally, in Changsha and Xiamen dialects, the surface form in the metrically weak position of each underlying tone was uniform, regardless of tonal contexts (see Figures 4, 7). Besides, the underlying T1 and T4 in Chengdu dialect were also free from the contextual effect, realized as [33] and [22] respectively. The contextual dependency in the four dialects could be summarized as (3).

On the contrary, the metrically strong units generally maintained their underlying pitch forms. This indicates they might have two morae that could bear contour tones. It should be noted that there were two exceptions, namely the T2/T3 in Chengdu dialect and T3 in Changsha dialect. In Chengdu dialect, the tonal alternations in σ(1) showed that T2 ([31]) and T3 ([53]) underwent [31] → [33] and [53] → [45], respectively. Many other factors might influence the surface form of lexical tones. For example, it has been generally deemed that the tonal merger (or simplification) is mainstream in tonal development in contemporary Chinese dialects (Pan, 1982). One possibility is that the exceptions in Chengdu dialects were triggered by tonal mergers, and that all surface tones of T3-TX approached T1([35])-TX. This kind of tonal merger could be supported by the case of Fuzhou dialect. In Fuzhou dialect, the surface tones of two syllables among T1-TX, T4-TX, and T5-TX were identical (see Table 5), indicating a completed tonal merger. However, the reasons why underlying tones changed in σ(1) of T2 in Chengdu dialect and T3 in Changsha dialect are still unknown and need more future studies to find out.

In a nutshell, despite a few exceptions, the four Chinese dialects we investigated exhibited consistent tonal representations of a level pitch contour in the metrically weak position while the underlying pitch form in a strong position. Therefore, this finding corroborated many case studies across Chinese dialects (Yue-Hashimoto, 1987; Chen, 2000), and the above tonal representations were generally confirmed by the fine-grained analysis (GCA) as we predicted (H3). Furthermore, this surface tone of the metrically weak position under the same metrical structure might be classified into context-independent and context-dependent types. These different types also supported our prediction (H2) that the specific pitch realization might be not identical within a certain metical pattern.

To date, although the tone sandhi in the metrically weak unit has been termed as the “metrical tone sandhi” (Zeng and Niu, 2006), many puzzles are still lingering. For instance, it is still unclear whether metrically tonal alternation is related to the duration pattern. We also wonder whether underlying tone would be realized as different surface forms according to different prosodic units. In the present study, we examined four Chinese dialects by drawing on the phonetic parameters of both duration and pitch, which offered a valuable chance to analyze their interaction. The “metrical tone sandhi” among the four Chinese dialects would be discussed below.

First, in Changsha dialect, the metrical structure belongs to the strong-weak pattern, exemplified by the shorter duration and surface level tones in σ(2). In terms of duration, the duration ratio in Changsha disyllabic words (1.53: 1) is close to that in Mandarin neutral-toned words (1.70: 1; Chen and Xu, 2006). We could assume that the σ(2) in Changsha dialect is monomoraic like the neutral-toned syllable in Mandarin (Duanmu, 1993), while the σ(1) is the heavy syllable containing two morae. Based on this pattern [σ(μμ) σ(μ)], the surface pitch realization of underlying T1, T2, T4, and T6 could be realized as (4) with T1 as the preceding context. Thus, the bimoraic rime of σ(1) in Changsha dialect could bear the underlying contour, while the surface tone of σ(2) could be realized as the level tone probably due to the limited capacity of tone bearing.

As for the disyllabic words in Chengdu dialect, the duration between two syllables is evenly arranged. The metrical pattern might be a syllabic trochee [σ(s) σ(w)]. Like Changsha dialect, the pitch contour of σ(2) in Chengdu dialect also undergoes a similar phonetic reduction, resulting in a level tone as the surface representation. For instance, the surface tone of T1 and T4 in σ(2) could be described as (5) when preceded by T1. Although the duration is not a phonetic correlate of the underlying metrical structure in Chengdu dialect, the strong-weak metrical pattern could be manifested by this tonal alternation in σ(2) (Qin, 2012).

Then, for Fuzhou dialect, the duration in σ(2) is significantly longer than that in σ(1), indicating that σ(1) and σ(2) belong to the light and heavy syllables, respectively [σ(μ) σ(μμ)]. Under this circumstance, the underlying T1, T2, T4, and T5 in σ(1) lose their original tone features and are assimilated by the following tonal contexts. These tonal alternations could be illustrated as (6) with T1 as the following tonal context. The tonal processes of tone deletion in σ(1) and leftward tone spreading from σ(2) could be seen in Fuzhou dialect.

Similarly, the duration in σ(2) is also longer than the σ(1) counterpart in Xiamen dialect. Thus, the syllable weight between two syllables in Xiamen dialect is different. Given this pattern of syllable weight [σ(μ) σ(μμ)], the tonal neutralization occurs in the σ(1) in terms of underlying T1, T2, and T3. When these lexical tones are followed by T1, the tone sandhi could be seen as (7). The tonal neutralization at σ(1) could be attributed to the limited tone-bearing capacity in the monomoraic syllable. Based on the duration and pitch realizations, we might confirm the metrically weak-strong structure in Xiamen dialect.

From the cross-dialectal perspective, the tonal process in the left-dominant structure contains contour reduction, apart from rightward tone spreading (Zhang, 2007). In the right-dominant, both leftward tone spreading and tonal neutralization could occur in the surface tonal representation. Although the above tonal processes are manifested in different manners, the core driving force is the underlying metrical structure in these four dialects. That is, the metrically weak syllable undergoes tone sandhi, while the strong syllable, where the metrical stress lies, is fully realized as the underlying pitch form (Duanmu, 1993; Chen, 2000; Zeng and Niu, 2006). In general, “metrical tone sandhi” enlightens us re-think the interaction between duration and pitch realizations in Chinese dialects. More cross-dialectal research, however, is needed to explore the fundamental effect of metrical structure on the tone sandhi more clearly.

Meanwhile, it is noteworthy that not all the tone sandhi in the four Chinese dialects could be interpreted as the “metrical tone sandhi.” For instance, the pitch change of the σ(1) in Chengdu dialect and the T4 ([21] → [53]) in Xiamen dialect are beyond the phenomenon of the “metrical tone sandhi.” To some extent, it might be more appropriate to regard them as dialect-specific tonal alternations.