1.1. Nasal codas in Standard Mandarin

The canonical variety of Standard Mandarin spoken in northern mainland China is described as having two nasal codas, alveolar /n/ and velar /ŋ/. The lingual constriction of both nasals is occasionally described as reduced or absent, mainly after the low vowel /a/; velum lowering and the resulting nasalization persist in these cases (Xu, 1993; Chen, 2000; Duanmu, 2007, p. 24). Not all combinations of the five monophthongal vowel phonemes /i y u ə a/ and the two coda nasals occur; the minimally different combinations of vowels and nasal codas that do occur are shown in Table 1.

Despite having a phonemic contrast between alveolar and velar nasal codas, Standard Mandarin listeners show some difficulty distinguishing the nasal place contrast after non-low vowels, particularly /i/. Error rates for coda nasal place identification after /i/ in Standard Mandarin can be as high as 20% (Mou, 2006, p. 125); this is much higher than the error rate of 10% expected for native contrasts (Polka, 1991). Furthermore, in gating tasks, native Mandarin listeners can predict the place of articulation of an upcoming nasal 80 ms through /ə/ or /a/, but not /i/ (Mou, 2006). Following /i/, nasal tokens are increasingly judged to be /iŋ/ at each successive gate closer to the nasal coda past the vowel midpoint (Mou, 2006, p. 125). These results suggest that Mandarin listeners are unable to distinguish the sub-phonemic coarticulatory fronting of velar nasals in the context of /i/ from the retracted tongue position typical for native phonemic alveolar /n/ in this vowel context.

Acoustic investigations of Standard Mandarin provide a possible explanation for why nasal place perception is worse after /i/. In the non-high vowels /a/ and /ə/, a difference in monophthongal vowel quality co-occurs with nasal place difference: These vowels are more fronted before coda /n/ compared to /ŋ/, a pattern sometimes referred to as Rhyme Harmony (Duanmu, 2007). Rhyme Harmony results in a 250–500 Hz F2 frequency difference between the vowels in Standard Mandarin /an/ versus /aŋ/ (Shih, 1995; Chen, 2000; Mou, 2006), which likely contributes to the stability of the /n/ – /ŋ/ contrast after /a/. Similarly, the F2 frequency of the vowel in /əŋ/ is also 250–500 Hz lower than that of /ən/ (Chen, 2000; Mou, 2006), and the /n/ – /ŋ/ contrast after /ə/ is generally described as stable.

In contrast, the high vowel /i/ interacts differently and less consistently with nasal codas: The F2 of /i/ is not obviously affected by the place of the following nasal (Mou, 2006, p. 117). Some speakers of Standard Mandarin employ a strategy of schwa insertion in the nasal rhyme /iŋ/, producing it as [iə̯ŋ] (Mou, 2006; Duanmu, 2007). The inserted schwa has an F1 of about 500 Hz, which is higher than the F1 value of /i/ at roughly 300 Hz (Mou, 2006, p. 128). While this vowel quality difference may allow Standard Mandarin listeners to distinguish alveolar and velar nasal codas, it is not clear that schwa insertion is consistent across speakers. Duanmu describes schwa insertion as mandatory (2007, p. 58; see Table 1), but in an investigation by Mou (2006) three out of five speakers of Standard Mandarin failed to exhibit schwa insertion, producing [iŋ] rather than [iə̯ŋ] for /iŋ/. Thus, co-occurring vowel quality differences observed in /a/ and /ə/, but not /i/, are likely to facilitate successful discrimination of coda nasals.

Overall, both Rhyme Harmony and schwa insertion enhance the place contrast in coda nasals (Keyser & Stevens, 2006) when they follow the non-high vowels. Because not all speakers use schwa insertion as a strategy to differentiate /in/ and /iŋ/, it is likely more difficult for listeners to predict the upcoming coda nasal based on the vowel quality of /i/. Based on these results, some researchers have proposed a reduction in the /n/ – /ŋ/ contrast after /i/ in the mainland Standard Mandarin nasal inventory (Chen, 2000; Mou, 2006). The regional Standard Mandarin varieties considered in the following sections show contrast reduction not only in this context but also after other non-low vowels.

1.2. Coda nasal neutralization in Taiwan Mandarin

From the previous section we can see that there is some evidence that the nasal place contrast in Standard Mandarin is neutralized after /i/. In this section, we consider prior research on nasal codas in Taiwan Mandarin, a regional variety of Standard Mandarin that has been widely investigated. Like Standard Mandarin, Taiwan Mandarin exhibits the aforementioned tendency for F2 to be raised in the vowels /ə/ and /a/ when they are followed by an alveolar nasal [n] (Lin, 2002, p. 7; Lai, 2008, p. 165). However, some crucial differences exist between Taiwan Mandarin and mainland Standard Mandarin. Specifically, there is a consensus that speakers of Taiwan Mandarin typically merge the nasal codas following /i/ and merge variably after /ə/, but never after /a/ (Chen, 1991; Lin, 2002; Hsu & Tse, 2007; Fon, Hung, Huang, & Hsu, 2011; Chiu, Lu, Weng, Jin, Weng, & Wang, 2019).

Taiwan Mandarin speakers often also speak Taiwanese Min Nan, also known as Hokkien, a Min variety not mutually intelligible with Standard Mandarin. At least some of the specifics of the neutralization pattern in Taiwan Mandarin have been attributed to interference from Taiwanese Min Nan. Although Taiwanese Min Nan, unlike Standard Mandarin, has a three-way nasal place distinction in codas between /m/, /n/, and /ŋ/ (Chen & Guion-Anderson, 2011), Southern Taiwanese Min Nan allows /in/ but not /iŋ/. Seemingly as a consequence, Southern Taiwanese Mandarin speakers are often described as neutralizing the nasal codas to [n] following /i/, a pattern not observed in speakers of Northern Taiwanese Mandarin (Fon et al., 2011).

What remains controversial is the extent and outcome of the neutralization. Based on transcription data, there is near-total agreement in the literature that Taiwan Mandarin speakers neutralize the two coda nasals to alveolar [n] after /ə/ (Chen, 1991; Lin, 2002; Hsu & Tse, 2007; Fon et al., 2011). However, in a recent ultrasound study of nasal codas, Chiu, Lu, et al. (2019) report inter-speaker variation even in the context of /ə/. Four of nine participants in Chiu, Lu, et al.’s ultrasound study did not merge the nasals after /ə/; of the remaining five, three produced a merged [n] in stimuli containing both /ən/ and /əŋ/, but another two produced a merged [ŋ] in the same environment.

The outcome of nasal coda merger is especially unclear following the high vowel /i/. In perception experiments using an identification task, the nasal resulting from merger is sometimes identified as alveolar. For instance, the three annotators in Yang (2010) report perceiving an alveolar nasal after /i/. Note that these results are consistent with the perceptual bias reported for alveolar nasals in English listeners (Zee, 1981). In contrast are perceptual results from Lin (2002), who reports a bias toward perceiving a velar nasal /ŋ/ after the high front vowel /i/, and a bias toward perceiving an alveolar nasal /n/ after /ə/. The articulatory outcome of the nasal coda merger after /i/ is also similarly unclear. Fon et al. (2011) suggest that speakers from southern Taiwan merge coda nasals after /i/ to alveolar nasals, whereas speakers from northern Taiwan merge coda nasals in the same context to velar nasals. In contrast, Chiu, Lu, et al. (2019), who report the only articulatory data collected to date on the merger, suggest instead that /i/ conditions merger to /ŋ/.

Overall, Taiwan Mandarin listeners variably merge /n/ and /ŋ/, but the resulting nasal’s place of articulation is unclear. Perceptual identification experiments alternately suggest that the outcome of neutralization is either alveolar or velar, both of which are structure-preserving outcomes in the sense that they are identified with one of the segments that participates in the neutralization. However, in all these perceptual experiments, listeners were forced to choose between alveolar and velar labels. Chiu, Lu, et al. also do not explicitly compare merged and unmerged nasals to identify the former’s place of articulation. Thus, we cannot rule out the possibility that the place of the neutralized nasals in Taiwan Mandarin is neither alveolar nor velar, but intermediate between the two, and not structure-preserving in the aforementioned sense. This could account for listener groups identifying the nasals sometimes as alveolar or velar in identification tasks. This brings us to the case of nasal coda merger in Shanghai Mandarin.

1.3. Coda nasals in Shanghai Mandarin and Shanghainese

Shanghai Mandarin is a Standard Mandarin variety spoken by residents of the Shanghai area which is often a second language learned after Shanghainese during schooling or acquired simultaneously with Shanghainese at home. Shanghainese is the mutually unintelligible local variety of Wu Chinese spoken in the Shanghai area. Unlike most mainland Standard Mandarin varieties, but like Taiwan Mandarin, Shanghai Mandarin is also reported to neutralize /n/ and /ŋ/ after the non-low vowels /i/ and /ə/ (Luo, 2015; Guan, 2019). In the limited research on Shanghai Mandarin, the neutralized post-/i/ nasals are also described as alveolar [n] (Guan, 2019; Luo, 2015). Both of these perceptual studies use an identification task; recall that based on results from such tasks, Taiwan Mandarin listeners have been reported to have either an alveolar or velar response bias when it comes to discriminating /n/ and /ŋ/ after /i/ (Mou, 2006, p. 125; Lin, 2002).

Coda nasal merger in Shanghai Mandarin has been explicitly attributed to the influence of Shanghainese (Luo, 2015), which lacks a coda nasal place contrast. Effects of the first language (L1) on incomplete acquisition of sound contrasts in a second language are ubiquitous; here we summarize findings that are specific to nasal codas. Increased experience with Chinese varieties which lack a coda nasal contrast has been shown to reduce listeners’ perceptual sensitivity to the coda /n/-/ŋ/ contrast in English (Wu et al., 2016) and Standard Chinese (Wang et al., 2018). Likewise, Burmese speakers who learned Mandarin as an L2 are less sensitive in discriminating Mandarin coda nasals (Lai, 2008), in keeping with Burmese having only non-contrastive coda nasality like Shanghainese (Green, 2005). Given the absence of a coda nasal place contrast in Shanghainese, it is then not particularly surprising that Shanghai Mandarin speakers have difficulty distinguishing some codas differing in nasal place. What cannot be explained by an L1 interference account is why mergers in Shanghai Mandarin are reported in some but not all vowel contexts.

An L1 interference account is also limited in its ability to predict the place of articulation of the neutralized nasals in Shanghai Mandarin. A logical possibility is that the neutralized nasals in Shanghai Mandarin have the same place of articulation as nasals in similar rhymes in Shanghainese. However, other than after the low vowels /a/ and /ɑ/, where lingual closure during nasal codas is largely agreed to be incomplete or inconsistent, descriptions of Shanghainese offer multiple candidates for the place of articulation of the nasal, sometimes varying by vowel context (Table 2). There is general agreement that the back non-high vowel /o/ is followed by a velar nasal (Sherard, 1972; Shen, 1981; Qian 2003, p. 56). But there is significant disagreement on how the nasal coda is articulated after the remaining non-low, non-back vowels. Some impressionistic accounts state that alveolar [n] occurs after /i/, /y/, and /ə/ (Shen, 1981; Xu & Tang, 1988), and Ping (2005) provides electropalatographic evidence that at least some tokens of the Shanghainese coda nasal in these environments are alveolar [n] (21). Qian (2003, p. 56) posits a palatal nasal coda [ɲ] for this group of rhymes, and still other analyses suggest a velar nasal [ŋ] (Sherard, 1972; Chen & Gussenhoven, 2015). Note that such variation in phonetic description of non-contrastive features, owing to the perceptual biases of transcribers, is a well-known limitation when relying solely on auditory transcription (Kerswill & Wright, 1990; Pouplier & Hardcastle, 2005; Munson, Edwards, Schellinger, Beckman, & Meyer, 2010). Regardless, such variability limits the utility of an L1 interference account. That is, these descriptions provide us with viable candidates for the merged place of articulation of nasals in Shanghai Mandarin under an L1 interference account, but no way to arbitrate between them.

1.4. The current experiments

In this paper we examined the extent of coda nasal place neutralization in Shanghai Mandarin, its outcome, and the vowel contexts which encourage it. We were motivated to investigate place neutralization because it has been studied less than neutralization of other features. The extant phonetic research on neutralization is heavily focused on voicing neutralization in coda obstruents (see Warner, Jongman, Sereno, & Kemps, 2004; Kleber, John, & Harrington, 2010; Roettger, Winter, Grawunder, Kirby, & Grice, 2014; Nicenboim, Roettger, & Vasishth, 2018) and flapping in North American English (see Herd, Jongman, & Sereno, 2010; Braver, 2014). These are usually described as structure-preserving processes in phonological accounts: That is, neutralization results in a sound structure present outside of the neutralizing environment, for instance a voiceless stop in the case of voicing neutralization. Instrumental phonetic studies of voicing neutralization demonstrate, however, that phonetic traces of the intended forms typically remain, making neutralization incomplete (Fourakis & Iverson, 1984; Port & O’Dell, 1985; Port & Crawford, 1989; Warner et al., 2004; Kleber et al., 2010; Roettger et al., 2014). Place neutralization, along with most other types of neutralization, has not been subject to this level of scrutiny (see Yu, 2011). It remains to be seen whether the phonetics of place neutralization are similar to those of voicing neutralization in generating novel phonetic structures from the participating segments.

Examining place neutralization conditioned by segmental context also provides an opportunity to probe contributions of perceptual and biomechanical biases to the process of contrast reduction. If outcomes are modulated by segmental context, as may be the case in Shanghai Mandarin, neutralization can be assessed with respect to substantive properties of this context; in the case of Shanghai Mandarin, for instance, neutralization may prove to be mediated by perceptual or articulatory properties of the preceding vowels. Unconditioned merger in place of articulation (Bukmaier, Harrington, & Kleber, 2014; Chang & Shih, 2015; Jannedy & Weirich, 2017; Chiu, Wei, Noguchi, & Yamane, 2019) does not afford this opportunity because it applies regardless of segmental context.

In our experiments, we first used ultrasound tongue imaging to examine nasal coda neutralization in production by Shanghai Mandarin and Standard Mandarin control speakers. We then played stimuli uttered during the production experiment to a separate group of Shanghai Mandarin and Standard Mandarin speaking controls. In this perception experiment, we used an AXB task instead of an identification task as has been done previously. Recall that forcing subjects to choose between alveolar or velar labels is not useful in cases where the place of articulation of the resulting merger may be neither alveolar nor velar. We used the perception results not to identify the place of the resulting neutralized nasals, but to probe the extent of neutralization in Standard versus Shanghai Mandarin. Comparisons of this sort are typically not made within the same study, even though studies of both Standard Mandarin (Mou, 2006) and regionally accented Mandarin (Yang, 2010; Chiu et al, 2019; Luo, 2015; Guan, 2019) suggest similar patterns of neutralization.

2.1. Participants

Five self-identified native speakers of Standard Mandarin (3 male, 2 female, ages 21–23, mean age = 22) and fifteen self-identified bilingual speakers of Shanghai Mandarin and Shanghainese (4 male, 11 female, ages 19–25, mean age = 21.87) were recruited as participants in the ultrasound experiment. All participants were students at the University of California, Los Angeles and thus also spoke English. Mandarin control subjects were confirmed to have no ability in Shanghainese and never to have lived in a predominantly Shanghainese-speaking region. All bilingual participants except for speaker 9 acquired Shanghainese and Mandarin simultaneously from birth; Speaker 9 acquired Mandarin at age five. All participants were offered $15 as compensation.

2.2. Materials

Stimuli consisted of 16 pairs of Mandarin disyllabic words (32 words in total) differing only in whether they ended with an alveolar nasal coda /n/ or velar nasal coda /ŋ/. Each pair of stimuli was roughly frequency-matched based on occurrence in the SUBTLEX corpus (Cai & Brysbaert, 2010). The stimuli included four preceding vowel contexts for the nasal codas at issue: /i/, /ə/, /a/, and the diphthong /ua/ (Table 3, see Table A1 in the Appendix for full stimuli). Stimuli were presented to both participant groups using the same simplified Chinese characters.

2.3. Procedure and equipment

Stimuli consisted of 16 pairs of Mandarin disyllabic words (32 words in total) differing only in whether their second syllable ends in an alveolar or velar nasal. Simultaneous ultrasound tongue imaging and audio were recorded in a sound booth in the UCLA Phonetics Laboratory. Ultrasound images were recorded using a Telemed Micro device outfitted with a Telemed MC4-2R20S-3 convex probe recording at a rate of 82 frames per second. The probe’s full field of view of 104° and an imaging depth of 90mm were used in all recordings. The probe was stabilized under each speaker’s chin with an Articulate Instruments UltraFit headset (Spreafico, Pucher, & Matosova, 2018). Audio was recorded using a Røde smartLav+ omnidirectional lavalier microphone attached to the headset approximately five centimeters to the right of the speaker’s mouth (sampling rate 44.1 kHz). Audio recordings were digitized using a Focusrite Scarlett 2i2 audio interface, which was also configured to receive the synchronization pulse train emitted by the ultrasound device as the time-aligned second channel of a stereo file.

Stimuli were presented as simplified Chinese characters. To prevent incorrect readings of the stimuli, subjects were briefly familiarized with the wordlist and frame sentence before recording. The experiment was presented using OpenSesame experimental software (Mathôt, Schreij, & Theeuwes, 2012). The stimuli were presented in a different pseudo-random order in ten blocks, for a total of 320 tokens for each speaker. Stimuli were embedded in a frame sentence, which was also presented in simplified Chinese characters (我知道怎么读 ___ Wǒ zhīdào zěnme dú ___ ‘I know how to read ___.’); participants in each of the two groups were instructed to produce the resulting full sentence with either a Shanghai Mandarin or Standard Mandarin reading respectively. All testing was carried out by the second author, a Standard Mandarin speaker who does not speak Shanghainese. The stimuli were located in utterance-final position in order to avoid influence on the production of the nasal codas of nasal place assimilation to, or resyllabification into, a following syllable. The experiment was self-paced: Participants started and ended the ultrasound and audio recording process using keypresses. The entire recording session lasted approximately 40 minutes.

2.4. Analysis

Audio files were segmented with the Montreal Forced Aligner (McAuliffe, Socolof, Mihuc, Wagner, & Sonderegger, 2017). The resulting segment boundaries were inspected in Praat (Boersma & Weenink, 2019) and hand-corrected as required. For the 40 tokens recorded for each vowel-nasal combination, spectrograms were inspected to confirm the presence of a lingual closure, signaled by lowered intensity during the force-aligned /n/ or /ŋ/ interval. Tokens whose nasal codas did not demonstrate lingual closure were discarded so as not to present a confound to the analysis of lingual closure place. Nasal codas in the /i/ and /ə/ contexts demonstrated a relatively consistent lingual closure, and most tokens were retained (average number retained per speaker: 38.95 /in/, 38.5 /iŋ/, 37.35 /ən/, 37 /əŋ/). Tokens following the vowel /a/ were more frequently excluded (average retained: 22.9 /an/, 26.6 /aŋ/), consistent with prior data on both Shanghainese and Standard Mandarin demonstrating a tendency toward reduced or absent lingual constriction after low vowels (Xu, 1993; Chen, 2000; Duanmu, 2007, p. 24). Tokens following /ua/ were retained somewhat more often (32.3 /uan/, 35.4 /uaŋ/). Out of the 320 utterances per speaker collected, an average of 269 tokens per speaker remained for analysis.

In order to evaluate the tongue shapes each speaker used to produce the two nasal codas, a single representative frame at the midpoint of each stimulus-final nasal coda token was selected based on the force-aligned segment boundaries and the synchronization signal emitted by the ultrasound device. After applying a series of filtering operations to reduce speckle noise in the ultrasound signal, which improves the signal-to-noise ratio for the analysis (Carignan, 2014), an arc-shaped region of interest was selected for each frame (Figure 1a–b) and the set of midpoint frames for each speaker was submitted to a separate principal components analysis (PCA), which takes each frame’s pixels as its numerical input.

Figure 1

Sample data for speaker 10; anterior is right. (a) Raw tokens of /n/ (left) and /ŋ/ (right); (b) the same filtered and trimmed to region of interest. (c) PC1 eigentongue for full image set, showing much of the variation between pixel values for /n/ frames (high PC1 scores, red) and /ŋ/ frames (low PC1 scores, deep blue). (d) Reconstruction of typical tongue position in /n/ and /ŋ/ frames from eigentongues for PCs 1–10 (see Section 2.5).

The resulting principal components (PCs) capture patterns of covariation among the pixels in the set of frames. These patterns can be displayed as so-called eigentongues (Hueber et al., 2007), maps relating the new dimensions of variation to scores assigned to individual frames. The eigentongue shown in Figure 1c, for instance, captures much of the covariation in pixel values across one speaker’s /n/ and /ŋ/ frames: PC1 scores positively correlate with brighter pixels in the red area and negatively correlate with brighter pixels in the deep blue area; these red and blue pixel clusters tend to have brighter values during /n/ frames and /ŋ/ frames, respectively (Figure 1d).

The first ten PCs were retained as a lower-dimensional representation of tongue shape for an individual speaker, explaining an average of 80.12% of variation (standard deviation = 2.57%). For each speaker, the scores for these PCs and the target of each token, i.e., velar /ŋ/ or alveolar /n/, were submitted to a linear discriminant analysis (LDA). The resulting linear discriminant (LD) score, when normalized to a [0,1] range for all speakers, may be taken as an index of how distinctly /n/- or /ŋ/-like the coda nasal in each token is: The LDA was structured such that /n/ was consistently near 0, and /ŋ/ was consistently near 1, for all speakers. We refer readers to Hueber et al. (2007) and Hoole and Pouplier (2017) for more information on the application of principal components analysis to image data, and to Mielke, Carignan, and Thomas (2017), Strycharczuk and Sebregts (2018), and Kochetov, Faytak, and Nara (2019) for applications of similar dimensionality reduction procedures to ultrasound data.

Statistical comparisons across vowel context and/or across speaker groups were done using linear mixed-effects models using lme4 (Bates, Mächler, Bolker, & Walker, 2014) in R v3.5.3 (R Core Team, 2019). These models maximize statistical power despite an unequal number of observations under each condition; they also allowed us to evaluate the contribution of the independent variables while controlling for baseline differences across speakers and items using the highest random effects structure with intercepts and slopes that converged. We then present reconstructions based on individual speakers’ first ten PCs’ eigentongues to verify the specific tongue shapes employed in the coda nasal contrast.

2.5. Ultrasound results

To determine whether there was a merger in the production of Shanghai Mandarin /n/ and /ŋ/, we compared the classification accuracy of each speaker’s LDA. The classification outcome for each item submitted to the LDA, on the basis of the linear discriminant function for each speaker’s data, was compared to the actual category membership, and thus used as a measure of predictability of the coda nasals in production given their expected pronunciation as velar or nasal.

Classification accuracies for speaker-specific LDAs, pooled for all speakers (Figure 2), were in keeping with previous findings that the frequency of merger varies across vowel contexts. Accurate classification of nasal codas following /a/ and /ua/ was at or near ceiling for all speakers, suggesting very distinct means of production across both groups; this is in line with previous findings that the contrast is maintained after low vowels. Accurate classification of nasal codas following the non-low vowels /i/ and /ə/ was also at ceiling for control speakers, demonstrating that a contrast is maintained in production by Standard Mandarin speakers across all four vowel environments tested.

Figure 2

LDA classification performance on each speaker’s basis data for nasal codas overall (left) and by rhyme (right).

Classification accuracy for bilingual Shanghai Mandarin speakers was considerably lower for rhymes containing one of the non-low vowels /i/ or /ə/, and accuracy for the rhymes /əŋ/, /in/, and /iŋ/ was often at or below chance. To explore further, a mixed-effects linear regression was run to predict LDA accuracy, with fixed effects of speaker group, vowel, coda nasal, as well as their interactions and random intercepts for subjects. There was a significant positive intercept, close to 1, indicating that LDA accuracy was overall quite high (β = 0.99; SE = 0.05; t(133.3) = 19.8, p < 0.0001). No main effects of group, vowel, or coda nasal on LDA accuracy reached significance, but there was a significant three-way interaction between speaker group, vowel, and coda nasal [χ²(10) = 102, p < 0.0001]. This interaction was driven by the Shanghai Mandarin group’s less distinct production of nasal codas only in the context of non-low vowels. Post-hoc comparisons confirmed that compared to Standard Mandarin speakers, Shanghai Mandarin speakers were less accurate at producing the velar nasal in the /ə/ vowel context [t(133) = 8; p < 0.0001], and both alveolar [t(133) = 5.2; p = 0.0001] and velar nasals [t(133) = 7.7; p < 0.0001] in the /i/ vowel context. In sum, the LDA accuracy data show that alveolar and velar nasals produced by the Shanghai Mandarin group were less distinguishable than those produced by the Standard Mandarin group in the /ə/ and /i/ vowel context.

The LDA performance data considered so far only reveal patterns in distinctiveness or similarity of articulations used for the coda nasals. To appreciate the actual tongue configurations employed, we first examined the overall distribution of LD scores by segment and vowel context. We use the preceding low nuclei /a/ and /ua/ as a point of reference since they were consistently well discriminated, suggesting distinct articulatory configurations. Since lenition disproportionately resulted in discarded tokens of the rhymes /(u)an/ and /(u)aŋ/, we pool the /a/ and /ua/ data in the analysis that follows to even out the sample sizes. We view this as appropriate since inspection of the LD scores revealed no difference in coda nasal articulation conditioned by preceding /a/ and /ua/ nuclei. Figures comparing LD score distributions for /a/ and /ua/ are given in the appendix (Figures A1–A2).

Using mixed effects regression, we modeled the LD value as a function of speaker group (Standard or Shanghai Mandarin), intended nasal (alveolar or velar), and vowel context (/a/, /ə/, and /i/). We also included random intercepts for speaker and item. There was a significant three-way interaction between speaker group, intended nasal, and vowel context, as expected based on the LDA accuracy scores [χ²(7) = 2643, p < 0.0001]. Planned comparisons confirmed that Standard Mandarin control speakers used distinct tongue configurations for intended /n/ and /ŋ/ in all vowel contexts [/a/: z-ratio = –30.783; p < 0.0001; /ə/: z-ratio = –23.677; p < 0.0001; /i/: z-ratio = –22.433; p < 0.0001]. Shanghai Mandarin speakers also used distinct tongue configurations after /a/ and /ə/, but not after /i/ [/a/: z-ratio = –28.112; p < 0.0001; /ə/: z-ratio = –8.011; p < 0.0001; /i/: z-ratio = –2.899; p = 0.1410]. Below, we discuss the distributional differences in tongue configurations for intended alveolar and velar nasals in more detail.

The distribution of linear discriminant (LD) values by segment and speaker group is shown in Figure 3 for the /(u)a/, /ə/, and /i/ vowel contexts. For the Standard Mandarin control group, regardless of preceding vowel context, alveolar nasals almost exclusively exhibited an LD value lower than 0.4 (98.5% of /n/ tokens), and velar nasals an LD value higher than 0.6 (99.7% of /ŋ/ tokens). We will henceforth refer to these as canonical ranges for the alveolar and velar nasals. The probability density distributions in the LD value for both nasals were compact and well-separated. The Shanghai Mandarin group had similar but less compact distributions for its coda nasals when preceded by /a/ (Figure 3, top right), with most alveolar nasals preceded by /a/ scoring lower than 0.4 (97.3% of /n/ tokens) and most velar nasals scoring higher than 0.6 (96.7% of /ŋ/ tokens). However, the distribution of LD values varied substantially from speaker to speaker for both of the Shanghai Mandarin group’s coda nasals when preceded by /i/ or /ə/, with the overall result that many fewer canonical nasals were produced (58.5% of /n/ tokens below 0.4 and 32.1% of /ŋ/ tokens above 0.6).

Figure 3

Distribution of LD values for nasal codas in the /(u)a/, /ə/, and /i/ contexts for the Standard Mandarin (left) and Shanghai Mandarin (right) groups.

Following /ə/, Shanghai Mandarin speakers’ LD values (Figure 3, middle right) showed evidence of inconsistent coda nasal neutralization to /n/, a result mainly driven by inter-speaker variation between three major patterns shown in Figure 4 (center). Two speakers showed control-like separation of /n/ and /ŋ/ after /ə/ (speakers 12 and 13). Five speakers had a near-complete overlap of the coda nasals after /ə/ (speakers 1, 3, 4, 6, and 19). The remaining speakers exhibited a more complex, bimodal distribution of the coda nasal in /əŋ/, which alternated between canonical /n/-like and canonical /ŋ/-like variants (i.e., speakers 15, 16, 17). Note that this pattern cannot be predicted by age of acquisition: All aforementioned speakers learned Standard Mandarin and Shanghainese simultaneously at home.

Figure 4

Distribution of LD values for nasal codas in the /(u)a/, /ə/, and /i/ contexts for each individual in the Standard Mandarin (top) and Shanghai Mandarin (bottom) groups. Trivially low probability density values (below 0.01) are not plotted for readability.

At the group level, this yielded a roughly unimodal, canonical /n/-like distribution for the /ən/ rhyme, with occasional /ŋ/-like productions contributing to a long tail approaching the /ŋ/ end of the LD scale. For /əŋ/, which had /n/-like characteristics much more often and in more speakers, there was an overall bimodal distribution containing /n/-like and /ŋ/-like peaks. The /ŋ/-like tokens were partly produced by non-merging speakers, but many were produced by speakers who variably produce the coda in /əŋ/ as /n/-like or /ŋ/-like. This distribution was driven in part by a bimodal distribution for all /əŋ/ items with the exception of 解梦 /t͡ɕiɛ˨˩.məŋ˥˨/ ‘dream reading’ in which speakers nearly always produce a canonical coda [ŋ].

Following /i/, LD values for Shanghai Mandarin speakers’ nasal codas suggest a more complete merger (Figure 3, bottom right). Most speakers did not distinguish between /n/ and /ŋ/ in production after /i/, as shown by the predominant pattern for individual speakers (Figure 4, right): a distinct cluster in the middle range of the LD value, flanked on either side by the canonical velar and alveolar nasals preceded by /(u)a/. This single non-canonical variant was typically centered around an LD value of about 0.4. While for many speakers these two distributions overlapped very closely (i.e., speakers 1, 13, and 19), others show more token-to-token variability in production, particularly for /iŋ/. Speakers 2 and 20, for instance, produced tokens of /iŋ/ ranging across nearly the entire LD range. This group generally exhibited two clear modes for both /n/ and /ŋ/ which fell just outside of the 0.4–0.6 range for non-canonical variants, but which nonetheless differed substantially from the modes for post-/a/ vowels. A small group of speakers (i.e., Speakers 2, 17, and 20) distinguished the two post-/i/ coda nasals to some extent, but there was considerable overlap between the categories, and some speakers (i.e., Speakers 15 and 16) tended toward a bimodal distribution for /iŋ/ with a small subset of tokens that were produced like canonical /ŋ/. This pattern does not appear to be driven by lexical factors as is the case for /əŋ/ above, since it affects all /iŋ/ lexemes. Rather, unclear mediating factors keep some /iŋ/ tokens distinct.

All in all, these data show that most Shanghai Mandarin speakers produced a single non-canonical variant after /i/. The remainder of this section is given over to identifying these non-canonical variants, the identity of which cannot be deduced from LD values alone: The LD encodes only similarity to the two maximally distinct articulations at either end, with the middle of the LD scale not necessarily representing an interpolation between these two variants. Anything not particularly similar to either maximally distinct nasal stop will fall in between; in this case, such segments could be nasal stops with full closure at a mid-palatal location, but could also potentially be reduced coda nasals which do not achieve full closure.

We used the PC scores which serve as the basis data for the LDA to determine the specific tongue shapes used for the nasal coda of each rhyme. In doing so, we also aim to rule out a contribution of coda nasal lenition or deletion to intermediate LD values, though efforts were taken to remove tokens of lenited nasals from inclusion in the analysis. Reconstructed typical ultrasound frames were created using the average value of each of the first ten PC scores from the PCA described in Section 2.4. The ten eigentongues, or maps of covariance in the original signal’s pixels described by each PC, obtained from the same analyses were multiplied by the resulting ten average PC scores and summed to obtain a graphical representation of the typical production for both nasal codas when preceded by the rhymes /i, ə, a, ua/. We used tongue reconstruction for this analysis because it is fully automated; in contrast, the more typical semi-automated process of tongue surface contour extraction and spline fitting requires manual annotator correction for accurate extraction of tongue surfaces from noisy ultrasound data. Such interventions are time-consuming and have been reported to decrease replicability (see Hoole & Pouplier, 2017; Roettger, 2019).

The resulting reconstructions (Figure 5) show displacements of the tongue dorsum and blade (for /ŋ/ and /n/, respectively) consistent with lingual closure during the nasal codas. As expected from the LD score data, Standard Mandarin control speakers employed distinct tongue shapes which align well with the intended place of the coda nasal (Figure 5, top; see Figure A3 for additional speakers). Investigation of the typical tongue shapes used by speakers in the Shanghai Mandarin group suggests that the non-canonical, merged nasal codas used by most of this group after /i/ were in fact articulated with tongue shapes intermediate in place between the alveolar and velar nasals (Figure 5, bottom; see Figure A4 for additional speakers). The mergers observed for Shanghai Mandarin participants in the LD data can be seen in the reconstructions: The two columns which contain tokens produced after /i/ and /ə/ (center, right) contain appreciably similar shapes, resembling /n/ in the case of the /ə/ context and a fronted velar in the /i/ context.

Figure 5

Mean reconstructed ultrasound frames at midpoint of coda nasals for a typical Standard Mandarin control subject (top) and Shanghai Mandarin subject (bottom). Anterior is right. Estimated pixel values below zero, which are artifacts of reconstruction, are flattened to zero. Note the overlap of estimated contours for the Shanghai Mandarin speaker’s /n/ (red) and /ŋ/ (blue) in the /ə/ and /i/ context.

Darker shading in the reconstructed frames indicates that the tongue surface is more consistently present at that location. The speakers shown in Figure 5 have a compact, unimodal distribution of productions for /n/ as well as /ŋ/, which is reflected by the clear reconstructions for each coda-vowel pair. Faint tongue contours which reflect variable production should not be confused with faint artifacts of reconstruction appearing at the edges of some frames (i.e., arcs at about y = 100 in Figure 5, top). Shanghai Mandarin speakers with more variable productions (i.e., multimodal distribution of LD value, particularly for /ŋ/) show fainter or more diffuse reconstructions; see Figure A4 in the appendix for examples.

2.6. Interim discussion

The ultrasound data presented above provide a nuanced position on coda nasal neutralization in Shanghai Mandarin. There are striking differences in both the outcome and consistency of neutralization among the vowel environments examined here. In keeping with previous literature, nasal stops before /a/ and /ua/ were articulated with two distinct tongue shapes, even by Shanghai Mandarin speakers. Nasals after /ə/ were articulated distinctly by Standard Mandarin control speakers and by some in the Shanghai Mandarin group. However, many in the Shanghai Mandarin group produced intended /ŋ/ in this environment as either canonical [ŋ] or canonical [n], with little in the way of intermediate articulations. This is evident from the bimodal distribution of velar nasals after /ə/ on the LD. This effect was asymmetrical: The LD value distribution of /ŋ/ was bimodal, typically merging to [n] but sporadically produced as [ŋ]; /n/ however, was largely unimodal and produced as [n]. These findings replicate and extend Chiu et al.’s (2019) findings of inconsistent merger after schwa in Taiwan Mandarin.

Nasals after /i/ presented the most consistent evidence of neutralization: Unlike post-/ə/ tokens, nasal productions by the Shanghai Mandarin group after /i/ were unimodally distributed and centered in the non-canonical middle of the LD scale. These non-canonical variants were produced by most Shanghai Mandarin speakers and, upon reconstruction, did not closely resemble either coda [n] or coda [ŋ] after low vowels, a context that does not promote neutralization. Our findings in the /i/ vowel context differ from Chiu, Lu, et al.’s (2019), who report velar nasals as an outcome of merger after /i/.

We argue that the merged non-canonical nasals are fronted velar or palatal nasals, fronted strongly enough as to be difficult to classify as canonical velars on the LD. This tongue configuration is in keeping with the occasional description of coda nasalization after /i/ in Shanghainese as palatal (Qian, 2003). If so, it may be due to transfer from Shanghainese to L2 Mandarin, since this degree of vowel-to-nasal coarticulation after /i/ is not observed in the control Standard Mandarin group. However, a more parsimonious account is that in Shanghai Mandarin this configuration is simply a result of coarticulation with the high vowel /i/. A similar explanation based on articulatory ease has been proposed for Taiwan Mandarin in Chiu, Lu, et al. (2019).

In the following sections, we made use of materials produced by speakers in the production study to conduct a follow-up perception study for nasal codas after /i/, in light of known perceptual difficulties in distinguishing the nasal codas in this context (i.e., Mou, 2006). We restricted the follow-up perception experiment described below to the /i/ context because after /ə/, any variation present for speakers resembled a substitution error of [n] for intended [ŋ], the rate of which varied among lexemes. Neutralization after /i/ was observed more consistently across all items, and thus likely creates more general perceptual difficulties.

3.1. Participants

Fourteen Standard Mandarin listeners (7 males, 7 females, mean age = 24.5) and 14 Shanghai Mandarin listeners (3 males, 11 females, mean age = 29.1) were recruited for the perception experiment. Listeners were compensated with course credit or cash payment ($12 USD). No participants reported a history of speech or hearing disorders. All participants filled out a questionnaire for their residential history and language background. None of the Standard Mandarin participants had experience with Shanghainese. All Shanghai Mandarin bilingual listeners self-rated their level of Shanghainese ability as native or native-like (at a minimum of 6 on a 7-point Likert scale). All listeners in the Shanghai Mandarin bilingual group acquired Shanghainese and Mandarin simultaneously in China except for listener 3, who acquired Mandarin at age 5, and listener 13, who acquired Mandarin at age 18 in the United States. The pattern of results reported later remains the same even when we exclude these two participants.

3.2. Stimuli

Stimuli were drawn from the audio recordings collected in the production experiment described above. Only nasal codas that followed the vowel /i/ were selected because they demonstrated the most consistent pattern where intended /ŋ/ fell in the intermediate range described previously as non-canonical (0.4–0.6). Target syllables were extracted from the original recordings and trimmed to exclude the syllable onset, which varies from token to token (see Table 3 and Table A1 in Appendix). Stimuli were normalized to an average amplitude of 70 dB.

Canonical and non-canonical /iŋ/ and canonical /in/ were selected for Shanghai Mandarin speakers based on LD value. Many Shanghai Mandarin speakers produced both canonical and non-canonical variants of the velar nasal after /i/, often varying in which variant was used across repetitions of a single word. In the remainder of this section, we use /N/ as a shorthand for non-canonical variants of velar nasals. Intended /iŋ/ with an LD value above 0.6 were treated as canonical /iŋ/, and intended /iŋ/ with an LD value between 0.4 and 0.6 were treated as non-canonical /iN/. Nearly all /in/ were canonical; canonical /in/ were selected with an LD value below 0.4. One canonical /in/ token produced by Standard Mandarin control speaker 08 and one canonical /iŋ/ produced by Standard Mandarin control speaker 18 served as anchor files in positions A and B in the AXB task. Twenty Standard Mandarin /in/ tokens and 20 Standard Mandarin /iŋ/ tokens (all canonical) produced by control speaker 7 were used as controls in experimental position X. Other test stimuli were drawn from four Shanghainese speakers (speakers 2, 16, 17, 20; 2 females, 2 males) who produced the full range of canonical and non-canonical nasal coda types described above. All three types of stimuli were drawn from each speaker’s materials: 14 canonical /in/, 7 canonical /iŋ/, and 7 non-canonical /iN/. This led to a total of 112 (56 canonical /in/, 28 canonical /iŋ/, and 28 non-canonical /iN/) Shanghai Mandarin test stimuli.

The acoustic characteristics of the vowels preceding nasals in Standard Mandarin and Shanghai Mandarin are summarized in Tables 4 and 5, respectively. As shown in Table 4, Standard Mandarin alveolar and velar nasals could potentially be distinguished by the duration of the preceding vowel, because the force-aligned vowels were longer before alveolars compared to velars; furthermore, as expected, F3 at vowel midpoint was lower for velars compared to alveolars. Note that the vowel duration differences in the perception stimuli in Standard Mandarin were not consistent with schwa insertion before velars, which would be expected to increase rhyme duration. Before velars, F1 was also higher at the vowel endpoint compared to the vowel midpoint, the opposite of what might be expected with schwa insertion (Mou, 2006). Table 5, in contrast, shows that neither preceding vowel duration nor formant frequencies at vowel midpoint or endpoint distinguished upcoming alveolar or velar nasal place in Shanghai Mandarin stimuli.

3.3. Procedure

The performance of each listener was evaluated using an AXB discrimination task presented using MATLAB. Each participant heard three syllables in each AXB trial, each separated by 500 ms. They were informed that these syllables were produced by different talkers. The first and third syllables were always distinct: one a canonical /n/ and the other a canonical /ŋ/. Participants were asked to determine if the middle syllable was more similar to the first or third syllable by clicking buttons “choice 1” or “choice 2.” They were asked to disregard the differences in speakers, vowels, and tones and instead to focus on the differences between the final consonants in each item. No feedback was provided. The experiment took place in a sound-treated room in the UCLA Phonetics Lab. Stimuli were presented through 3M PELTOR (HTB79A-02) closed-system over-ear headphones. The experiment began with ten practice trials to familiarize subjects with the task. In the test phase that followed, participants were presented 304 trials in one sitting. These trials included counterbalancing of 152 files by switching the order of anchor files in position A and B, which contained one canonical /n/ and one canonical /ŋ/ produced by two Standard Mandarin control speakers. The presentation of trials was randomized. The experiment took about 40 minutes.

3.4. Analysis

To determine whether Shanghai Mandarin and Standard Mandarin listeners were able to distinguish coda nasals, we used mixed-effects logistic regression (Bates et al., 2014) in R v3.5.3 (R Core Team, 2019). The model included the between-subjects variable, listener group (Standard Mandarin or Shanghai Mandarin), the within-subjects variable, stimulus type (Standard Mandarin, Shanghai Mandarin canonical, or Shanghai Mandarin non-canonical), and their interaction. Standard Mandarin alveolar nasals and Standard Mandarin listeners were coded as reference conditions. Finally, random intercepts for individual speakers of stimuli and individual listeners were also included. This was the maximal random effects structure with which the model converged. Fixed effects were evaluated by comparing the log likelihood of two models with and without the variable of interest using a Likelihood Ratio test in R. Follow-up planned comparisons were conducted using emmeans (Lenth, 2020).

To determine how well Shanghai Mandarin and Standard Mandarin listeners were able to distinguish the different stimulus types, we used d-prime. Values were calculated separately for each listener group using R v3.5.3, for a total of six groupings by stimulus type (Standard Mandarin, Shanghai Mandarin canonical, and Shanghai Mandarin non-canonical) and listener type (Standard Mandarin, Shanghai Mandarin).

3.5. Perception results

The full logistic regression model had a significant negative intercept (β = –0.8; SE = 0.32; z-value = –2.5, p = 0.01), indicating that all listeners had a bias towards categorizing nasals as Standard Mandarin /ŋ/. Additionally, the interaction of listener group and stimulus type was significant [χ²(4) = 33.98, p < 0.001].

The effects plot for the interaction term is presented in Figure 6. Planned comparisons confirmed that Standard Mandarin listeners were able to distinguish Standard Mandarin /n/ from Standard Mandarin /ŋ/ [z-ratio = –9.1; p < 0.0001], Shanghai Mandarin /n/ [z-ratio = –4.4; p = 0.0005] and Shanghai Mandarin canonical /ŋ/ [z-ratio = –4.5; p = 0.0003] and non-canonical /N/ [z-ratio = –4.3; p = 0.0007]. But they failed to distinguish between Standard Mandarin /ŋ/, Shanghai Mandarin /n/, and Shanghai Mandarin /N/. That is, Standard Mandarin listeners were only able to distinguish between Standard Mandarin /n/ and canonical /ŋ/, and categorized all Shanghai Mandarin nasals the same as Standard Mandarin /ŋ/.

Figure 6

Effects plot for mixed-effects logistic regression. Standard Mandarin /n/ and the Standard Mandarin listener group served as reference levels.

In contrast, Shanghai Mandarin listeners failed to distinguish Standard Mandarin /n/ from any other stimulus type [Standard Mandarin /ŋ/: z-ratio = –2.2; p = 0.5; Shanghai Mandarin /n/: z-ratio = –2.8; p = 0.1; Shanghai Mandarin canonical /ŋ/: z-ratio = –2.2; p = 0.5; Shanghai Mandarin noncanonical /N/: z-ratio = –2.25; p = 0.4]. That is, Shanghai Mandarin listeners failed to distinguish between /n/ and /ŋ/ produced by either Standard Mandarin or Shanghai Mandarin speakers.

Furthermore, we can see from their d-prime values that even Standard Mandarin listeners were not too successful in distinguishing Standard Mandarin /n/ and /ŋ/ (Figure 7). Native-like performance typically results in d-prime values of 3; Standard Mandarin listeners had average d-prime values of 0.2. As expected from the logistic regression results reported previously, d-prime values were close to zero for Shanghai Mandarin nasals; their d-prime value for the Standard Mandarin /n/-/ŋ/ contrast was only marginally higher.

Figure 7

d-prime for AXB discrimination task by talker and stimulus type.

In summary, we see from the perception results that Shanghai Mandarin listeners failed to distinguish both Standard Mandarin and Shanghai Mandarin /in/-/iŋ/, providing evidence for neutralization of coda nasal place after /i/ in Shanghai Mandarin. Standard Mandarin listeners also failed to distinguish Shanghai Mandarin /in/-/iŋ/, regardless of the similarity of the articulatory configuration to canonical productions. Additionally, although they were able to distinguish Standard Mandarin /in/-/iŋ/, the near zero d-prime value attests to a near merger in Standard Mandarin as well.