What R Mandarin Chinese /ɹ/s? – acoustic and articulatory features of Mandarin Chinese rhotics

1.1. Rhotics in Mandarin Chinese

Different phonetic symbols have been used in the literature to represent Mandarin rhotic sounds, which will be introduced in detail in the next two sections. For ease of discussion, in the present study we will use the symbol for post-alveolar approximant /ɹ/ to represent rhotics in prevocalic and postvocalic positions, and the symbol of the syllabic approximant /ɹ̩/ in syllable-nucleus position, to indicate that they are all rhotic sounds.

1.2. The current study

This study explores the variability in the phonetic realization of rhotics in various syllable positions in Mandarin Chinese. We aim to answer three research questions. The first question concerns the articulatory features of Mandarin Chinese rhotics. As reviewed in previous sections, mixed results have been found regarding the tongue shapes of Mandarin rhotics. It would be interesting to see if tongue-shape variation – tip-up retroflex and tip-down bunched tongue shapes – can be found in the post-alveolar rhotic in Mandarin Chinese, as such variation has been reported in post-alveolar rhotics in other languages. Ultrasound evidence will be used to categorize the various tongue shapes used in Mandarin Chinese.

The second aim of this study concerns the effects of syllable position and vowel context on the phonetic characteristics of rhotics. Mandarin Chinese rhotics in different syllable positions have different characteristics. The prevocalic rhotic is treated as a consonant (Duanmu 2007; Lin 2007; Lee and Zee 2003; Lin and Wang 2013), while the postvocalic rhotic in the r-suffix is considered either as part of a (rhotic) vowel (Duanmu 2007; Hu 2020; Lee and Zee 2003) or as a consonant (Chao 1968; Lin 1989, 2007; Wang 1993). It is necessary to investigate the effect of syllable position and the effect of preceding and following vowels on the acoustic and articulatory features of Mandarin Chinese rhotics.

The third aim of the study relates to the frication noise in Mandarin Chinese rhotics. The existence of frication noise in the prevocalic position is largely unexplored. Examining the frication noise helps us understand the nature of the rhotic sounds. Therefore, the third strand of the research will determine the frequency of rhotic realizations that have measurable frication.

By providing a systematic phonetic description of Mandarin Chinese rhotics, the current study can add further phonetic evidence to the literature on rhotics in the world’s languages. It can also provide phonetic evidence that may help us resolve the debate as to the precise nature of the prevocalic rhotic.

2.1. Participants

Eighteen native Mandarin speakers (4 males and 14 females) from Northern China were recorded using ultrasound imaging. As mentioned in Section 1.2.2, r-suffixation (er-hua) is a common feature of Mandarin Chinese and Mandarin dialects spoken in various places in Northern China, such as Beijing, Shandong Province and Hebei Province. All the speakers in the current study were born and grew up in these three places. Besides their Mandarin dialects, they could all speak fluent Mandarin Chinese and used r-suffixation naturally in daily communication. Their average age was 23.3 years old (Range: 21–-28, SD = 1.99) at the time of the experiment. They were postgraduate students studying in Hong Kong and the United States and also spoke English as an L2. Their English proficiency ranged from IELTS 6.5 to 8.5. The details of their English proficiency are reported in Chen et al. (2024).

2.2. Stimuli

The stimuli included words containing the prevocalic rhotic followed by the [ʅ a ɤ u] vowels, the postvocalic rhotic preceded by the [i ɿ ʅ y u a ɤ] vowels, and the syllabic rhotic (see Appendix). The Mandarin Chinese low vowel /a/ has three allophones, [a], [ɑ] and [ɛ] (Lin 2007). Due to phonotactic constraints, the prevocalic rhotic cannot be combined with [a] and [ɑ] without a coda. Thus, [ɹan] and [ɹɑŋ] were used in the current study. The tone of the tested words was not controlled for because there are some accidental gaps. For example, all words with the prevocalic rhotic bear Tone 4 except [ɹan₃₅] “but” 然 which has Tone 2. This is because [ɹan] cannot carry Tone 4. Sonorant-initial syllables exhibit a strong tendency to co-occur with Tone 2, a pattern that is plausibly the result of diachronic sound changes. Furthermore, in Mandarin Chinese, not every word can undergo r-suffixation. By consulting native speakers, we first ensured that words with the postvocalic rhotic were natural and frequently used in daily communication, rather than keeping the tone consistent in different vowel contexts. Therefore the stimuli with the postvocalic rhotic in different vowel contexts had different tones (Tone 1, Tone 2, Tone 3). For words with the syllabic rhotic, we included three tones (Tone 2, Tone 3, Tone 4). This allowed post hoc comparisons examining the effect of tone on tongue shape. There are no words with the syllabic rhotic and Tone 1, and thus they were not included.

There are some commonly used minimal pairs consisting of syllabic and postvocalic /ɹ/s in Mandarin Chinese. These minimal pairs are two-character words that are identical in segmental combination, but the syllable position of the rhotic is different. One is a compound word where the rhotic is syllabic and the word is pronounced as two syllables (such as /y₃₅.ɹ̩₂₁₄/ 鱼饵‘fish bait’), and the other is a word with r-suffixation where the two syllables are pronounced as only one syllable (such as /yɹ₅₅/ 鱼儿 ‘small fish’). In /y₃₅.ɹ̩₂₁₄/, the rhotic sound is syllabic, while in /yɹ₅₅/, the rhotic sound is a coda (r-suffix). Five such minimal pairs were included (see Appendix) to see if syllable position and syllable boundary would affect the articulatory and acoustic features of the rhotics.

The Mandarin Chinese words were produced in the carrier phrase /tʂɤ₅₁ kɤ ___ pa/ “This is ___.” The carrier phrases were designed to have as little a coarticulatory effect as possible. The target words were embedded between the mid vowel /ɤ/ and the bilabial stop /p/. The word /pa/ is a sentence final particle with a neutral tone. The [p] sound does not have any lingual target and the /a/ is pronounced as [ə] because of its particle status, so it should have a lesser coarticulatory effect on the target word. The word /tʂɤ₅₁ kɤ/ “this” in the carrier phrase is a function word, so the vowel [ɤ] is reduced and its phonetic realization is also close to a schwa [ə]. The reduced vowel quality of [kə] is very different from content words such as [ɤ₃₅] ‘goose’ that have a full vowel. Therefore the carrier phrases have very little coarticulatory influence on the target words. All Mandarin Chinese prompts were presented in Chinese characters. We consulted native speakers to make sure that the words used did not have multiple pronunciations. All stimuli were randomized and repeated eight times.

2.3. Procedure

Before the experiment, all the participants were briefed about the experimental procedure and ultrasound machine and signed the consent forms. They were also asked to read through the stimulus list to familiarize themselves with the words. During the experiment, participants were seated in a soundproof booth, facing a computer screen that displayed the prompts. At the beginning of each session, speakers were asked to swallow a sip of water to make the hard palate visible, following Stone (2005). They were then asked to raise their tongue tip to touch the alveolar ridge, and then move the tongue tip back along the midline of their mouth as much as possible. These two actions were used to capture the ultrasound image of the hard palate. All speakers repeated the two actions multiple times until an image of the hard palate was clearly captured.

2.4. Ultrasound data acquisition

Two ultrasound imaging systems with the same stimuli and experimental procedure were used in this experiment. One was the Siemens ACUSON X300 ultrasound system at Haskins Laboratories with blue dots head correction (Chen et al. 2017; Noiray et al. 2020; Whalen et al. 2005). The other system was an EchoB ultrasound machine together with the Articulate Assistant Advanced (AAA) software (Articulate Instruments Ltd 2012) at the Chinese University of Hong Kong. Of the 18 speakers, 6 were recorded with the Siemens system in the Haskins Laboratories, and 12 were recorded with the EchoB system at the Chinese University of Hong Kong. The compatibility of data collected with the two systems will be discussed below.

With the Siemens ACUSON X300 system, the ultrasound probe was held on a microphone stand and positioned under the participants’ chins during the experiment. The probe could move freely with the jaw. The participants were asked to look at the screen in front of them where the stimuli were presented. In order to image the midsagittal plane of the tongue, the experimenter stood in front of the participants and reminded them to avoid side-to-side head movements or rotation during the recording. The participants were asked to adjust their head position and repeat the words when there were any out-of-plane movements.

The relative position between the probe and the head was not constant. To make the ultrasound images comparable across frames, the ultrasound splines from the raw images had to be corrected according to the movements of the probe and the head. Two video cameras were positioned in front and at the side of the participants to record the front and side views of the participants’ faces in order to get head movement information during recording. The head movement was represented by the movement of blue dots on the participants’ heads. The blue dots’ movement was tracked by a tracking algorithm implemented by the in-house MATLAB procedure DotsTracking.

For bunched gestures, the frame where the gesture reached the maximum constriction in the post-alveolar region was selected as the representative frame. For retroflex gestures, the frame with an additional bright line above the tongue surface was selected. There were usually one or two frames containing a bright white line in the retroflex data. If there was more than one frame with a bright white line, the frame where the bright line was closest to the post-alveolar region was selected. On the representative ultrasound frame, the tongue splines were drawn with the interactive MATLAB procedure ‘GetContours’ (Tiede 2018). The tongue splines were exported as 100 equally spaced data points from ‘GetContours’ for head movement correction. The articulatory data were collected at a frame rate of 36 frame/s.

Before the tongue spline correction, synchronization of the data from the ultrasound machine with that from two video cameras was achieved through cross-correlation, a method that measures the temporal displacement between signals. This process enabled the calculation of the time lags between the ultrasound and the cameras. The head movement was then corrected according to the blue dot positions implemented by an optimization method implemented in MATLAB (Chen et al. 2017). The correction algorithm moved and overlapped the dots as much as possible and exported head-corrected tongue splines.

With the Echo B system, the articulatory and acoustic data were collected with the Articulate Assistant Advanced (AAA) software. The ultrasound probe was stabilized under the chin with a headset made by Articulate Instruments Ltd. to make sure that the relative position of the probe and the head was largely maintained (Articulate Instruments Ltd 2008). The software recorded ultrasound videos and audio signal, and automatically synchronized the two signals. The ultrasound videos were recorded at a frame rate of 60 frame/s. The synchronized ultrasound videos were segmented and labeled manually in AAA. A key frame where the maximum constriction could be seen was selected as the target frame of typical rhotics. The tongue splines in the key frames were manually tracked, with the aid of the “autofit” function in AAA that could automatically smooth the splines based on the ultrasound images. Each spline was exported as 124 equally spaced data points.

The splines were drawn on the lower boundary of the lighter line which represents the tongue-air interface in the ultrasound images. Figure 1a demonstrates the raw ultrasound image of a bunched tongue gesture, and the fitted splines on the ultrasound images. Due to the limitations of ultrasound imaging, it is hard to draw tongue splines of retroflex tongue shapes when part of the tongue front is not imaged well. In the current study, to illustrate the tongue shapes based on all the useful information in the ultrasound images, only the visible part of the tongue in the retroflex tongue shape was drawn, as exemplified in the right panel of Figure 1b. The position of the tongue front and tongue tip, therefore, cannot be seen in the fitted splines.

The major differences between the Siemens and Echo B systems are the frame rates and stabilization methods. The ultrasound videos collected from the Siemens system have a frame rate of 36 frame/s, while videos from the Echo B system have a frame rate of 60 frame/s. A higher frame rate means that the ultrasound machine captures more ultrasound images per second. When producing approximants, tongue movements are relatively slow, and 30 frame/s has been shown to be sufficient to capture them (Lawson et al. 2011; Mielke et al. 2016). A faster frame rate is needed for sounds like flaps, clicks and stops because the tongue moves faster when producing them (Stone 2005). In the current study, the temporal resolution of both ultrasound systems was sufficient for the purpose of examining the tongue movements of rhotics. The higher frame rate in the Echo B system resulted in some repeated ultrasound images of the same tongue shape. As for the stabilization techniques, both methods have been proven to be efficient in maintaining the relative position between the ultrasound probe and the head, or correcting for such movements (Chen et al. 2017; Scobbie et al. 2008). Therefore, differences in frame rate and stabilization method do not influence the reliability of the data. The splines from the two ultrasound systems were never compared in one statistical model. Therefore, although two systems have been used to acquire the ultrasound data, the data from the two systems are comparable for the purposes of this study, and caution has been taken to make sure that the data analysis is legitimate.

2.5. Ultrasound data analysis

The raw ultrasound data were first visually inspected and described. The tongue shapes were then categorized as bunched or retroflex based on the tongue tip position. We classify tongue shape into these two broad types for three reasons. First, there is a controversy in previous literature regarding whether the production of Mandarin rhotics involves curling or pointing up of the tongue tip. Bunched and retroflex tongue shapes differ mainly in tongue-tip position. Therefore categorizing tongue shape into these two types can properly address the debate. Second, the two-way categorization based on tongue tip position is more practical than other categorizations. Different methods of categorization have been adopted in previous studies to describe Mandarin rhotics and retroflex consonants. King and Liu (2017) categorized the postvocalic rhotic into three categories – tip up, front up and front bunched. Xing (2021) categorized rhotic tongue shapes into two broad categories (retroflex and post-alveolar) with five sub-types (curled up, tip up, front up, flat post-alveolar, domed post-alveolar). They included curled up, tip up and front up in the retroflex category. Luo (2020) used a three-way categorization (bunched, retroflex, and humped tongue shape) for retroflex consonants. Although Jiang et al. (2019b) did not categorize the tongue shapes of rhotics directly, they suggested that the tongue shapes of the syllabic and postvocalic rhotics in Northeastern Mandarin are reminiscent of the bunched and retroflex contrast of the English /ɹ/. From a practical perspective, dividing tongue shapes into many categories with slight differences could possibly lead to a decrease in reliability of classification and a lower inter-rater consistency when there are multiple raters because the tongue shapes (especially the different bunched tongue shapes) can be quite ambiguous. It would also be less convenient for cross-study comparisons or cross-dialectal comparisons. Therefore, using a two-way categorization is a practical compromise. Third, the two-way categorization of rhotic sounds has been widely adopted by studies of other languages, such as English and Dutch (Delattre and Freeman 1968; Mielke et al. 2010, 2016; Scobbie and Sebregts 2010). English rhotics can be produced with various tongue shapes, and many different ways of categorizing English rhotics have been used (Delattre and Freeman 1968; King and Ferragne 2020; Lawson et al. 2013). But the two-way categorization is the most widely adopted one in the literature (Mielke et al. 2010, 2016; Twist et al. 2007; Zhou et al. 2008) and therefore is used in the current study.

Following previous studies, two basic criteria were used in the categorization: 1) which part of the tongue was used to make the constriction, and 2) whether the tongue tip was pointing up or pointing down (Mielke et al. 2010, 2016). While it was sometimes difficult to determine the position of the tongue tip based on a single ultrasound frame, a sequence of tongue contour movements between the segments before and after the rhotic was examined. The first author and another trained phonetician experienced in ultrasound imaging did the categorization. They initially worked separately, then discussed the different judgments together. If they had the same categorization, or they agreed with each other after their discussion, the judgment of that particular token was marked as “same”. If they disagreed with each other even after the discussion, the judgment was marked as “different” and that token was not analyzed. The inter-rater agreement for all tokens in this study was 93.98 %.

Generalized additive mixed models (GAMMs) were used to quantify the tongue curves tracked from the ultrasound data using the “mgcv” packages in R (Wood 2023). GAMMs have been used to model lingual movements or tongue splines in recent ultrasound studies (such as Heyne et al. 2019). We used a customized function to visualize ultrasound splines fitted by GAMMs using “itsadug”, “magrittr”, “plotly” and “stringr” packages. Polar coordinates were used to model the tongue contours because it has been proposed that tongue root position is better estimated with polar coordinates instead of Cartesian coordinates (Mielke 2015). The data points exported from “GetContours” and AAA were in Cartesian coordinates. They were converted into polar coordinates to conduct GAMMs with calculated origins for each speaker, and then converted back into Cartesian coordinates for plotting. According to Mielke (2015), the origin of the polar coordinates is an approximation of the center of an imaginary circle that corresponds to the arc of the tongue traces. For the Siemens system, the x value of the origin was the midpoint of the x values of all splines of the sample tokens for each speaker. The y value was chosen at a point where the connected lines between the origin and the tongue surface both at the root and at the tip were approaching perpendicular. For the Echo B system, the x coordinate of the origin was calculated using the formula: Depth* (1+(offset/Pixperscanline)) * cos (90-(FOV/2)) (angles in degrees), and the y coordinate of the origin was 0 (personal communication with Dr Alan Wrench).

2.6. Acoustic data acquisition and analysis

The audio recordings from the Siemens system were extracted from the ultrasound videos. They were segmented using forced alignment (FAVE-align) and then manually adjusted (Rosenfelder et al. 2011). The audio files from the EchoB system were exported from AAA and labeled manually. The rhotic sounds and their flanking vowels were labeled in PRAAT (version 6.0.36) (Boersma and Weenink 2017). For the prevocalic and postvocalic rhotics, we did not segment the rhotic sounds and the preceding/following vowels. The first three formants of the whole /ɹV/ or /Vɹ/ sequences were tracked at ten equidistant points. The formants were measured using linear predictive coding (LPC) in PRAAT, and the maximum formant was set as 5,000 Hz and 5,500 Hz for male and female speakers respectively. The formant values where F3 was the lowest were identified as the acoustic target of the rhotic sound, and were extracted by an R algorithm. The algorithm ensured that the minimum F3 was extracted from the first half of an /ɹV/ syllable or the second half of a /Vɹ/ syllable. The raw formant data were plotted and visually inspected to make sure that abnormal data were excluded. The first three formant values were then transformed into the Bark scale for further analysis.

To quantify the frication noise in the prevocalic rhotic, we measured the zero-crossing rate (ZCR) for all syllables containing this sound and the subsequent vowels. ZCR is defined as the number of times the speech signal crosses zero in 1 s, calculated by dividing the number of zero-crossings by the window length. It can serve as a reliable measure of frication noise intensity, where a higher ZCR indicates increased aperiodicity. Following the approach of Shao and Ridouane (2023), we obtained upward and downward zero-crossing points in a 40 ms sliding window and analyzed 50 data points in each syllable. We modeled the time-normalized ZCR in different vowel contexts using Generalized Additive Mixed Models (GAMMs).

The first three formants and the difference between F3 and F2 (F3-F2) of rhotics were examined with linear mixed-effect models using the lmer () function from the ‘lme4’ package (version 1.1–21) (Bates et al. 2015). In addition to examining the first three formants, F3-F2 was also examined because this measure can partially correct the differences in speakers’ vocal tracts, and thus is less influenced by individual differences in age, gender and height (McAllister and Tiede 2017). In English, the primary acoustic cues for the /ɹ/ sound are a low F3 and a small F3-F2 difference (Delattre and Freeman 1968; Lindau 1985; Westbury et al. 1998; Zhou et al. 2008). To examine the effect of syllable position, linear mixed-effects models were conducted on F1, F2, F3 and F3-F2 with Syllable position as fixed effects, Participant and Item as random effects (both Participant and Item as a random intercept, and Participant also as a random slope). The models with best fit are presented with p values calculated with the ‘lmerTest’ package (Kuznetsova et al. 2017), and post-hoc comparisons were done with ‘emmeans’ package (Lenth 2018). In addition to the static formant frequency of rhotics at the lowest F3, we also used GAMMs to analyze formant trajectories of /ɹ/ together with adjacent vowels to get a better understanding of how these sequences were coarticulated.

3.1. Articulatory features of Mandarin Chinese rhotics

3.2. Acoustic features of Mandarin Chinese rhotics

4.1. Articulatory features of Mandarin Chinese rhotics

Our articulatory and acoustic data show that Mandarin Chinese rhotics can be articulated with various tongue shapes in syllabic and postvocalic positions, and no significant difference was found between bunched and retroflex tongue shapes in the first three formants. This seems to suggest a many-to-one relationship between articulation and acoustics for Mandarin Chinese rhotics, which has also been observed for the English /ɹ/ (Delattre and Freeman 1968; Lindau 1985; Westbury et al. 1998).

Although the tongue shape variation in Mandarin Chinese rhotics is similar to the well-established bunched-retroflex variation in English and Dutch rhotics, our data showed that Mandarin Chinese rhotics have language-specific features in their articulation. First of all, vowel context does not categorically change the tongue shape of Mandarin Chinese rhotics. Mielke et al. (2010, 2016 suggested that in English, retroflex tongue shapes are preferred when the /ɹ/ sound is adjacent to low or back vowels due to compatibility of gestures. Retroflexion involves retraction of the tongue body, which is also required for English low back vowel production. The tongue shapes of bunched gestures involve the raising of the tongue front, which is similar to English high front vowels. Therefore, retroflex tongue shapes are more compatible with low back vowels, while bunched tongue shapes are more compatible with high front vowels in English. In Mandarin Chinese, however, tongue shape is not affected by vowel context. If a speaker chooses to use a bunched or retroflex gesture, he/she would continue with that tongue-shape category in spite of any segmental changes. One may argue that high vowels do not prevent retroflex tongue shapes in Mandarin Chinese because the postvocalic rhotic does not directly follow high front vowels. Phonological analysis suggests that a schwa is inserted between a high front vowel and postvocalic rhotic during r-suffixation (Lin 1989). The perceived insertion of schwa is due to the coarticulatory influence of the high front vowel on the subsequent rhotic sound. In spectrograms, the F3 gradually lowers when transiting from a high front vowel to the following /ɹ/ sound. In terms of articulation, the inserted schwa is the result of the tongue passing through a schwa-like configuration (Gick and Wilson 2006). Phonologically, in Mandarin Chinese, this phenomenon is interpreted as a rule-driven insertion of schwa. On a phonetic level, it represents a coarticulatory pattern that is comparable across both English and Mandarin Chinese languages. Hall and Hamann (2010) also suggest that the avoidance of the rhotic plus high front vocoid sequences is a cross-linguistic phenomenon likely grounded in articulatory phonetics.

One possible reason that lingual compatibility does not influence the production of Mandarin Chinese rhotics to the same extent as English rhotics might be because high vowels in Mandarin Chinese are not as high as English high vowels. The F1 of the English high front vowel /i/ is about 270 Hz and 310 Hz for male and female speakers (Peterson and Barney 1952), while the F1 of the Mandarin Chinese /i/ is about 300 Hz and 401 Hz for male and female speakers (Zee and Lee 2001). Apical vowels in Mandarin Chinese are more like mid vowels than high vowels in terms of their acoustics and articulation (Zee and Lee 2001; Lee-Kim 2014). The tongue shape of Mandarin Chinese rhotics, therefore, is affected by the raising of the tongue front to a lesser degree than in English, resulting in fewer conflicts in articulation. In other words, high vowels in Mandarin Chinese are not as elevated, allowing better compatibility with both tongue shapes. Consequently, there is no distinct preference for either bunched or retroflex tongue shapes, providing speakers with more flexibility in their choice of which to use.

A similar phenomenon has been reported in King and Liu (2017). They examined the Mandarin postvocalic rhotic of 12 Mandarin speakers using ultrasound imaging and found little coarticulatory effect of vowels on the tongue shape of the following rhotic. The cumulative results suggest that lingual compatibility may not be as important a factor in Mandarin as it is in English. Mandarin Chinese speakers can “overcome” the incompatibility between retroflex tongue shapes and high front vowels to use their preferred tongue shape in production. In syllable positions where articulatory variation is allowed (syllabic and postvocalic positions), individual speech production strategies determine tongue shape in the production of Mandarin Chinese rhotics. Another aspect worth noting is that all the tokens for each vowel context in each syllabic position were elicited by eight repetitions of one or two test items. We cannot entirely dismiss the possibility that the null effect of vowel context may stem from a lack of variability in the stimuli. In speech perception, previous studies on categorical perception have shown that stimulus repetition may introduce a bias in the experiment results (Rogers and Davis 2009). It is plausible that the repetition of stimuli could similarly influence production outcomes. Future investigations incorporating a broader range of materials would be valuable to validate this potential impact.

Another difference is the effect of syllable position. In English, tongue-shape variation can be found in all syllable positions, and retroflex tongue shapes are more frequently found in the prevocalic /ɹ/ than in other syllable positions (Heyne et al. 2020; King and Ferragne 2020; Mielke et al. 2010, 2016). Mielke et al. (2016) proposed that the preference for a larger anterior gesture in syllable-onset position might result in more instances of the retroflex /ɹ/ in prevocalic position. In the current study examining Mandarin Chinese, however, tongue-shape variation can only be observed in postvocalic and syllabic positions. The prevocalic rhotic was consistently found across speakers to be produced only with bunched tongue shapes. This seems to suggest that the prevocalic rhotic is different from the postvocalic and syllabic rhotics to some extent in terms of tongue-shape variation. As reviewed in Section 1.1.1, there has been a long debate on the phonological status of the prevocalic rhotic because of its special phonetic qualities. A detailed discussion on this issue based on our articulatory and acoustic data will be provided in Section 4.2. For the syllabic and postvocalic rhotics, where both bunched and retroflex tongue shapes can be found, syllable position does not affect tongue-shape categories. Each speaker consistently used the same tongue shape for syllabic and postvocalic rhotics. Unlike our study but also using ultrasound imaging, Xing (2021) reported tip-up retroflex tongue shapes in prevocalic position in 8 out of 18 Beijing speakers. One difference between the current study and Xing (2021) is the criteria for tongue shape categorization. Xing (2021) classified Mandarin Chinese tongue shapes in various syllable positions into three categories: retroflex, bunched and post-alveolar. Within the retroflex category, three subcategories were identified – curled up, tip up, and front up, while the post-alveolar category included two subcategories – flat post-alveolar and domed post-alveolar. In her examination of the prevocalic rhotic, Xing (2021) observed both retroflex and post-alveolar tongue shapes. In contrast, our current study employed a more conventional two-way categorization, relying on tongue tip position as the primary criterion. Specifically, if the tongue tip pointed upward, the tongue shape was categorized as retroflex; otherwise, it was classified as bunched. Consequently, in the present study, front-up tongue shapes from Xing (2021) were reclassified as bunched, which may have played a role in the differences observed between the two studies. Another possible reason for the discrepancies between the current study and Xing (2021) is that ultrasound imaging does not provide clear visualization of the tongue tip when the tongue is perpendicular to the probe. The poor imaging quality of the tongue tip could lead to controversy regarding its position. Another possible reason is that the current study included participants from Beijing, Hebei and Shandong, while the participants in Xing (2021) were all from Beijing. Although the participants all spoke Mandarin Chinese during the experiment, we cannot entirely exclude the possibility that their Mandarin dialects may have had some influence on tongue shape.

In addition, the percentages of bunched and retroflex tongue shapes are different in Mandarin Chinese and English. According to Mielke et al. (2010, 2016, bunched tongue shapes are much more prevalent in all syllable positions in English. Of the 27 American English speakers they examined, 16 used only bunched gestures, 2 used only retroflex gestures, and 9 used a mix of bunched and retroflex gestures. However, our data show that the occurrence of bunched and retroflex tongue shapes is about equal in postvocalic and syllabic positions for Mandarin Chinese rhotics (10 bunched speakers and 8 retroflex speakers).

In addition, we also found a more retracted tongue root in postvocalic rhotics compared to prevocalic rhotics. This is consistent with the results of Lee and Zee (2014) and Xing (2021). Pharyngeal retraction has been found in rhotics in many languages such as English (Delattre and Freeman 1968; Zhou et al. 2008), Upper Sorbian and Brazilian Portuguese (Howson 2018). Our data provides evidence for the suggestion in Howson (2018) that a constriction in the pharynx achieved by retracting the tongue root might be a shared characteristic of rhotics.

4.2. Nature of the Mandarin Chinese prevocalic rhotic

In the current study, we found that the Mandarin Chinese prevocalic rhotic can only be articulated with bunched tongue shapes, which is different from the articulatory variation observed in the syllabic and postvocalic rhotics. This finding is consistent with Lee (1999) and Zhu and Mok (2023). Lee (1999) examined the tongue shapes of Mandarin prevocalic rhotics in four speakers using palatograms and linguagrams, while Zhu and Mok (2023) conducted a study involving eight Beijing Mandarin speakers and eight Japanese-Mandarin simultaneous bilinguals using ultrasound imaging. Both studies did not report any instances of retroflexion of the tongue. The current study did not find retroflexion either, even with a greater number of speakers. As for the acoustic properties, the prevocalic rhotic is characterized by a low F3, around 2,321 Hz. Moreover, frication noise was frequently observed in various vowel contexts in prevocalic rhotics before a clear formant structure characterized by a low F3. Nevertheless, speakers who did not produce frication noise in prevocalic rhotics could also be found. In summary, the prevocalic rhotic is characterized by a low F3 and an optional fricated onset. The complexity of its phonetic properties might be the reason why the prevocalic rhotic has been an “enigmatic” sound which is hard to categorize. Based on acoustic and articulatory features, the prevocalic rhotic can be categorized as a post-alveolar approximant with an optional fricated onset or as a voiced post-alveolar fricative that is devoiced in some vowel contexts. We will discuss the two types of analyses and the evidence for them below. The two categories assume the same place of articulation, but differ in manner of articulation. Both involve the approximation of the tongue tip and the post-alveolar region, with the fricative having a narrower oral constriction.

The first account is that the prevocalic rhotic is underlyingly a post-alveolar approximant, but sometimes it can be realized phonetically with additional frication at the beginning. The two phonetic realizations are free variations. This account is supported by three kinds of evidence. First, the prevocalic rhotic can be found both with and without frication noise, suggesting that frication noise is not an obligatory phonetic component in the production of this sound. Both variants are perceptually acceptable for the prevocalic rhotic. Second, frication noise is more frequently observed in high vowel contexts than in low vowel contexts. It suggests that the frication noise might be the result of a relatively higher tongue position (either bunched or retroflex) in high vowel contexts compared to other vowel contexts and caused by a coarticulatory effect. The high vowel context is a contributing factor because of the narrowing between the tongue and the hard palate in terms of aerodynamics and airflow. This differs from gestural coarticulation, as observed in Mandarin Chinese rhotics, where tongue shapes remain unaffected by vowel context. Third, this sound is characterized by a low F3, which is the primary acoustic cue for the syllabic and postvocalic rhotics, and for the approximant /ɹ/ in English. Luo (2020) showed that the retroflex fricative /ʂ/ and retroflex affricates /tʂ tʂ^h/ did not feature a low F3. The mean F3 of /ʂ/ reported in her study was around 3,100 Hz, while the mean F3 of the prevocalic rhotic in our study was 2,321 Hz. It indicates that the prevocalic rhotic has a much lower F3 compared to the voiceless fricative /ʂ/. In addition, Hu (2020) examined Mandarin Chinese fricatives [s ʂ] and the prevocalic rhotic produced by four Beijing Mandarin speakers using electropalatography (EPG). The linguopalatal contact for both the prevocalic rhotic and [ʂ] was in the post-alveolar region, but the percentage of linguopalatal contact was smaller for the prevocalic rhotic. This provides direct evidence that the tongue makes less contact with the palate in the prevocalic rhotic than in a typical fricative with the same place of articulation, suggesting that [ʂ] is a post-alveolar fricative and the prevocalic rhotic is a post-alveolar approximant.

Another account is that the prevocalic rhotic is underlyingly a voiced fricative. This is because the prevocalic rhotic differs from the postvocalic and syllabic rhotics in both acoustics and articulation. In terms of articulation, tongue-shape variation cannot be found for the prevocalic rhotic. Though our data only represent 18 speakers from Northern China, this is consistent with previous studies using other articulatory measures (Lee 1999). The lack of variation in tongue shape suggests that the prevocalic rhotic is different from the syllabic and postvocalic rhotics, which are uncontroversially approximants. However, it needs to be noted that only 18 speakers in their twenties were tested in the current study. It is possible that retroflex tongue shapes might be found in prevocalic rhotics if more speakers, or speakers from other age groups, were examined.

In terms of acoustics, frication noise is only observed in the prevocalic rhotic and not in the syllabic or postvocalic rhotics. The lack of frication noise in some prevocalic tokens can be caused by devoicing. According to Ohala (1997), the production of voiced fricatives is more complex than that of voiceless fricatives due to aerodynamics. To produce frication noise, the oral air pressure has to be as high as possible in order to produce a large velocity of air through the oral constriction. In contrast, to produce voicing, one has to maximize the pressure difference across the glottis by keeping the oral air pressure as low as possible. Therefore, “fricatives favor voicelessness” (Ohala 1997). This is why voiced fricatives are less common than voiceless fricatives in the world’s languages (Ohala 1983, 1994). Consequently, a voiced fricative might sometimes lose the frication noise because speakers fail to maintain frication, as we observed in this study.

The lowered F3 might be caused by lip protrusion. It has been suggested that lip constriction can lower F3 by increasing the volume of the front cavity (Espy-Wilson et al. 2000). Using a rather small sample size (3 speakers) and older speakers, Tiede et al. (2019) also observed lip protrusion in the retroflex consonants [ʂ tʂ^h tʂ] and the prevocalic rhotic in Taiwan Mandarin. It is possible that the prevocalic rhotic involves a constriction at the lips which also causes a low F3. However, the existence of lip protrusion is not the only factor that lowers F3 so this does not necessarily lead to the conclusion that this sound is a fricative. The production of English approximant /ɹ/ also involves lip protrusion (Delattre and Freeman 1968; King and Ferragne 2020; Zhou et al. 2008). Furthermore, a closer examination of the formants of prevocalic rhotics in the current study shows that F2 was not lowered. It casts doubt on the lip protrusion account because lip protrusion should also lower F2, although it was observed that increased lip protrusion for the English /ɹ/ had an impact on F3 and not on F2 (King and Ferragne 2019). To further clarify the question, future studies are needed to examine the existence of lip protrusion and its influence on F3 lowering.

Although the data from the current study do not allow us to draw a firm conclusion, the approximant account seems to be more compatible with the phonetic cues of the prevocalic rhotic, while the fricative account cannot be completely ruled out.

From a phonological point of view, Natvig (2020) argues that the category “rhotic” in the phonological system is an unspecified sonorant consonant. The particular surface form of rhotic is a result of its relationship to other potential liquid phonemes and their phonological properties, as well as the computations to make all underspecified representations pronounceable. Sebregts (2015), in contrast, advocates for characterizing rhotic variants through “family relationships”. Sebregts (2015) argues for establishing diachronic links between different rhotic variants, exploring historical and geographical variations in the data. By closely examining the phonetic details of rhotics in a large corpus, particularly urban accent data, Sebregts (2015) aimed to identify the origins of specific variants, often tied to casual speech processes like lenition. Chabot (2019) proposed that the membership of rhotics is arbitrary. The cross-linguistic commonality of rhotics is procedural stability and diachronic stability. Procedural stability implies that rhotics implicated in phonological processes can vary phonetically without affecting the process itself, while diachronic stability suggests that the phonetics of rhotics can change over time without impacting their phonotactics.

In light of the perspectives outlined in these studies, it is evident that relying solely on the phonetic characteristics of Mandarin Chinese is insufficient to address the phonological status of rhotics. Investigations on diachronic and geographical variations are needed for a comprehensive understanding. Therefore, future studies could approach this question from the following perspectives to better resolve the debate. The first possibility is to check the historical development of words containing the prevocalic rhotic. The prevocalic rhotic in Mandarin originated from a nasal sound in Middle Chinese (Chinese spoken from the 6th to 10th centuries in the Southern, Northern, Sui, Tang, and Song dynasties), but it is less clear how it became oralized and gained frication (Karlgren 1915–1926; Hu 2020). The historical origins of the prevocalic rhotic seem to support the approximant account because changing from a nasal to an approximant is more natural than changing from a nasal to a fricative. Another possible line of inquiry is to examine how the prevocalic rhotic is realized in other Chinese dialects. A few studies have investigated the r-suffix in some Chinese dialects such as Liaoning dialect (Jiang et al. 2019b) and Hangzhou Wu (Yue and Hu 2019), but the prevocalic rhotic in other dialects is yet to be examined. Finally, examining this issue from a phonological perspective may bring further insights. Xing (2021) asked speakers from Beijing to produce a sequence of Vr#rV and found that speakers assimilated the prevocalic rhotic to the postvocalic rhotic in the preceding syllable. She proposed that Mandarin Chinese prevocalic and postvocalic rhotics share the same [(+)rhotic] feature.

Supplementary Material

This article contains supplementary material (https://doi.org/10.1515/phon-2023-0023).