An acoustic comparison of German tense and lax vowels produced by German native speakers and Mandarin Chinese learners Open Access

Much research has been done to demonstrate that Mandarin Chinese learners show phonetic inaccuracies in their production of English spoken as the second language (L2) (Chen, 2006; Chen et al., 2001). Although a handful of attempts have been made to examine the production of German vowels by Mandarin learners from a phonological or phonetic point of view (Fluck et al., 1984; Gao, 1984; Hachenberg, 2003; Hunold, 2009; Wang, 1988, 1993; Yen, 1992), few studies have been conducted from the acoustic perspective, despite the increasing number of L1 Mandarin speakers learning German. Among the sparse acoustic-phonetic studies, a too limited number of speakers (one German speaker and 11 Mandarin Chinese learners) were recruited in the investigation to generalize the findings (Ding et al., 2007). Thus, an extension of the previous work with more participants, better designed experiments, an optimized technique, and statistics is necessary.

The tense-lax opposition has been viewed as one of the most distinguishing features of the German vowel system. In German, without consideration of the schwas and /εː/, there are 14 vowel monophthongs that can be grouped into seven pairs, the members of which differ exclusively with respect to tenseness. The seven tense vowels are /aː oː eː iː uː yː øː/, whereas their lax counterparts are /a ɔ ε ɪ ʊ ʏ œ/. In contrast, the six monophthongs found in Mandarin Chinese /a o ɤ i u y/ are classified as tense vowels. The lax vowels do not occur in the Mandarin vowel system (Chen, 2006) and are considered to be difficult to learn for Mandarin speakers. It is generally acknowledged that tense and lax vowels in German are distinguished by both length and quality, although the spectral difference may be small, as for the German /aː-a/ pair. Thus, a line of studies has been conducted in terms of durational difference (Liu et al., 2014), or spectral difference (Chen et al., 2001; Jin and Liu, 2013), or both separately (Chen, 2006; Wang and Heuven, 2006) to illustrate how differently native speakers and L2 learners of English realize the tense-lax vowel contrast. However, the duration-quality relationship is mutually dependent and not the same for all vowels in German (Weiss, 1974). Therefore, it is very important to explore the duration-quality relationship as a whole. The current study aims to examine how Mandarin Chinese learners differ from native German speakers in producing the tense-lax contrast in terms of duration and quality as well as their interactions, so as to identify those German vowels that may be difficult for Mandarin learners and determine what kind of different acoustic cues Mandarin speakers employ in their production to distinguish German tense-lax vowels.

Two groups of young speakers were recruited for the study. The native group included 30 (15 females and 15 males) German native speakers (DEU group). They were German students studying at TU Dresden with a mean age of 23.6 years (range: 18–38). The L2 group included 30 (15 females and 15 males) Mandarin Chinese learners of German (CHN group). They were born in China and had no residency history in German-speaking countries before 18 years old. They were either students studying German as a major at Shanghai Jiao Tong University [having passed the examination of Prüfung für das Germanistik-Hauptstudium (PGH)] or students who had passed the German language examination (up to DSH-2 or DAF-16) before taking up their studies in German at TU Dresden. These L2 participants could be regarded as advanced German learners. They also had experience with English learning for three hours per week during their secondary school in China. Their average age was 24.1 years (range: 18–31). All participants, according to their self-reports, had normal speech and hearing functions with no history of any communication disorders, and they were financially compensated for their time in participation.

Each of the 14 German vowels was produced in monosyllabic nonsense words /dVt/. To indicate the tense-lax contrast orthographically in the words shown to the participants, the letter “h” was appended to indicate tense vowels, and a double-“t” was used to indicate lax vowels. Thus, these 14 monosyllables consisted of daht-datt, deht-dett, diht-ditt, döht-dött, doht-dott, düht-dütt, duht-dutt. Most of these were nonsense words, but the speakers had no difficulty in reading them with the expected vowels. In order to make sure that all the syllables were produced with similar prosody, the target syllable was further embedded in the carrier sentence “Ich habe /dVt/ gesagt (I have said /dVt/).” There are two main reasons for the employment of monosyllabic nonsense words. First, to focus on the formant and inherent duration differences between German native speakers and Mandarin Chinese learners in this study, we have to control the influences of other factors such as familiarity and orthography of the words or stress of the syllables, etc. Because real words do not allow this type of control, a frequent experiment paradigm is the use of nonsense words, which is popular in many other phonetics studies (Fischer-Jørgensen, 1990; Strange and Bohn, 1998). Second, all these /dVt/ pseudo-words are legal phoneme strings according to German phonotactic rules, syllables embedded in a carrier sentence could be spoken naturally by both German native and Mandarin L2 speakers. By randomizing each group of 14 sentences five times, we created a reading list of 70 sentences, thus making sure each vowel was produced five times with different intra-group orders by all speakers. All recordings took place in a studio equipped with a recording console (Behringer Eurorack MX1602). The studio microphone (Microtech Gefell M930) was placed at a distance of approximately 20 cm from the speaker's mouth. After a brief instruction and practice, the speaker was asked to read all of the 70 sentences. Each speaker was recorded separately and digitized at 44.1 kHz sampling rate and 16-bit resolution.

A total of 4200 utterances, each containing one target vowel, were recorded. For the phonetic annotation, the software webmaus (Kisler et al., 2017) was first employed to carry out a forced-alignment automatically, then the phoneme boundaries were manually adjusted by a phonetic expert following standard phonetic criteria (Peterson and Lehiste, 1960), taking into account both the acoustic landmarks (e.g., the first/last complete pulse period or the beginning/end of stable formant trajectories of vowels) and zero-crossing positions by referring to both visual and audio cues simultaneously.

For the formant measurement of vowels, most studies employed linear predictive coding (LPC) and the number of poles for the analysis was individually selected according to the age and gender of the speakers [e.g., Hillenbrand et al. (1995)]. However, the required LPC order also partly depends on the individual vowel. Therefore, we extended the procedure (Kathiresan et al., 2017) and implemented an automatic formant measurement method with the software praat (Boersma and Weenink, 2019) in our current study. Unlike their method dealing with the middle 0.3 s sound nucleus of each long vowel produced in a sustained style, we took into consideration the middle portion (from 20% to 80%) of each vowel. By using the LPC Burg algorithm in praat, with a frequency range of 0–5.5 kHz and a pole number of 10, 12, or 14, the first three formants (F1, F2, and F3) were extracted and averaged for the middle 60% portion of each vowel. Then these three groups of mean formant values measured with different settings for the pole number were further examined by referring to the table of each vowel's formant ranges of females and males. Only when the first three formants fell within the corresponding ranges was the whole group kept for further processing. Finally, we complemented formant ranges for six additional German vowels on the basis of Kathiresan et al. (2017) with the reference values of Pätzold and Simpson (1997) and our available data. In the last step, we calculated the root sum square (RSS) of standard deviation σ from the formant tracks according to Eq. (1). The mean formant values obtained for the pole number producing the lowest RSS were chosen as the final formant measurements of that vowel,

By using this strategy, we automatically measured the formants for 4068 vowel tokens in an accurate and reproducible way. The formants of 115 CHN tokens and 17 DEU tokens were manually checked because their extracted formant values fell out of the candidate formant ranges, which resulted in the failure of the automatic estimation. Considering the non-linear relationship between the formant frequencies and the corresponding perceived vowel quality, we further transformed the formant frequencies from the Hertz scale to the Bark scale. Finally, the values of all five repetitions of each vowel for a specific speaker were averaged to represent his/her production characteristic.

We compared the mean duration values and the F1-by-F2 vowel spaces for the 14 German vowels produced by DEU and CHN speakers in Secs. 3.1 and 3.2, respectively.

A proper vowel normalization was done in this study for three reasons: (a) to eliminate the speakers' anatomical/physiological information such as the gender-related difference, (b) to preserve the sociolinguistic information resulting from the speakers' regional background, (c) to obtain the phonemic information. Here, we transformed the Bark frequencies into Lobanov's z-scores. The procedure was individually applied to each formant (F1, F2, or F3) for each individual speaker. The formant values were transformed by subtracting that speaker's mean value calculated from all tokens, and then dividing the difference by that speaker's standard deviation. With this kind of normalization, the phonemic and sociolinguistic information could be best preserved and the anatomical/physiological characteristics could be attenuated, so as to enable the comparison of vowel spaces characterizing the DEU and CHN speaker groups across genders [see Adank et al. (2004) for a detailed overview of vowel normalization].

Since the 14 German monophthongs fall into seven tense-lax vowel pairs contrasted by the durational and spectral cues, we further examined the difference between two speaker groups with regard to the interaction between two acoustic cues in Sec. 3.3. For each vowel pair, we used two metrics to describe the tense-lax contrast produced by a specific group. The first one was the mean durational distance (DD) by computing the absolute durational difference between the tense and lax vowels for each speaker and then averaging across all speakers. The second one was the mean spectral distance (SD), which was calculated across all speakers within each group as shown in Eq. (2),

where the $F1¯, F2¯$⁠, and $F3¯$ were the mean formant frequencies (in Bark) of five repetitions for a vowel. The subscripts t and l indicated the tense and lax vowels in each tense-lax vowel pair produced by the individual speaker. N (=30) was the number of speakers within each group.

The mean duration for German vowels is plotted in Fig. 1 separately for tense (upper panel) and lax (lower panel) categories produced by DEU (left bars) and CHN (right bars) speakers. We conducted an unpaired t-test for each vowel produced by the two groups. For each of the seven lax vowels, the mean duration of the CHN group was significantly longer than that of the DEU group (at a significance level $α<0.01$⁠, marked with a dashed line between the left and right bars). For tense vowels, a significant difference between the DEU and CHN groups was only found for the high vowels /iː yː uː/ (at a significance level $α<0.05$⁠, marked with a dotted line between the left and right bars). An interesting phenomenon for tense vowels was that the mean duration of the CHN group was much longer than that of DEU group for high vowels (/iː yː uː/), and not much longer for mid vowels (/eː øː oː/), and even slightly shorter for low vowel (/aː/). To examine how speakers contrasted the duration in each pair, we further calculated the mean duration ratio of tense vowel to lax vowel (as shown between the upper and lower panels), which is a complementary metric to the durational distance in Sec. 3.3. Although CHN speakers produced generally longer durations than DEU speakers for all vowels with the exception of /aː/, the mean duration ratios of the DEU speakers were larger than those of CHN speakers. A t-test revealed that a significant difference between DEU and CHN groups was found in vowel pairs /yː-ʏ/, /eː-ε/, /øː-œ/, and /aː-a/.

The German vowel spaces produced by DEU and CHN speakers are illustrated in Fig. 2. Female and male speakers' data are jointly shown in the F1-by-F2 panel after the formant values were Bark-transformed and z-score normalized. As can be seen from Fig. 2, the tense vowels produced by the DEU group take up the peripheral area of the vowel space while the lax counterparts are more centralized. However, the German vowels produced by the CHN group show less spectral distinction, i.e., lax vowels are not more centralized than their tense counterparts, especially for the mid-front with lip-rounding vowel pairs (e.g., /øː-œ/ and /yː-ʏ/). Each pair of vowels produced by the CHN speakers is roughly located halfway between the tense-lax vowels produced by DEU speakers. For each vowel, we conducted the multivariate analysis of variance (MANOVA) with F1 and F2 as the dependent variables and speaker group as independent variable. The results showed significant differences of quality for all vowels (p < 0.001) except for the vowels /ε/ [$F(2,57)=0.63, p=0.54$⁠, Wilk's Lambda = 0.98] and /aː/ [$F(2,57)=2.43, p=0.097$⁠, Wilk's Lambda = 0.92].

The different realization of the tense-lax contrast of each vowel pair by DEU and CHN speakers can be found in Fig. 3. As lip-rounding is also a distinguishing feature in German, we incorporated the F3 (representing lip-rounding) into the spectral distance calculation, which is a supplement to Sec. 3.2. As we can see from Fig. 3, for each vowel pair, the circle (CHN) moves downwards and to the left (except for vowel pair /iː-ɪ/) from the square (DEU), indicating a shorter spectral distance (SD) and a shorter durational distance (DD) for CHN speakers compared with DEU ones. In other words, the tense-lax contrast produced by CHN speakers was less distinct than that by DEU speakers. For each vowel pair, we further carried out a t-test, with speaker group as independent variable and DD or SD as dependent variable to examine whether the realization of tense-lax was significantly different. The results demonstrated quite different relationships between the two speaker groups: for the DD, a significant difference was found only for the pair /aː-a/ [$t(58)=−2.70, p=0.009$]; however, for the SD, a non-significant difference was found only for the pair /eː-ε/ [$t(58)=−1.58, p=0.118$].

The current study was a production experiment including speech data from 60 speakers (balanced in number, gender, age, and the German language proficiency) for the analysis of the production of the German tense-lax vowel contrast. By employing improved techniques for extracting formant values and optimal normalization methods for analysis, we have illustrated that compared with German native speakers, Mandarin Chinese L2 learners employed different strategies to indicate the tense-lax contrast in terms of duration, quality, and duration-quality relationship. The main differences can be explained by the discrepancy between the Mandarin and German phonological and phonetic systems in terms of vowel space, tense-lax feature, and inverse quality-duration relationship.

The less distinctive spectral patterns in tense-lax German vowels for Mandarin learners may be attributed to the Mandarin Chinese vowel space. As we know, the six Mandarin monophthongs are spread in the peripheral areas, so tense vowels are relatively easier for Mandarin learners to acquire than lax vowels. As shown in Fig. 1, Mandarin speakers have achieved better performance for tense vowels than lax ones, which is consistent with the findings obtained in investigating English vowel production by Mandarin speakers (Chen, 2006; Chen et al., 2001). Besides lax vowels, lip-rounding front tense vowels are also located in the centralized vowel space, which are also difficult for Mandarin learners. Therefore, the most difficult vowel pairs for Mandarin learners to acquire are the front rounding ones /øː-œ/ and /yː-y/ (with no statistical significance in spectral distance), for their relative centralized locations are absent in Mandarin vowel space. Moreover, Mandarin speakers exhibit significant differences from German speakers in producing all lax vowels (except for /ε/ and /a/). There may be several reasons why Mandarin speakers can produce a similar lax vowel /ε/ as German natives. One reason is that the Mandarin diphthong /a$ɪ̂$/ is similar to the monophthong /ε/; another reason may be that their experience of learning English /ɛ/ may help them acquire this phonetically similar German vowel.

Apart from the different vowel space in Mandarin, Mandarin Chinese does not distinguish tense and lax vowels. Therefore, Mandarin speakers produce both tense and lax vowels halfway between tense-lax positions of German native speakers. This phenomenon is further confirmed by inspecting the tense-lax contrast in the interaction panel between durational distance and spectral distance in Fig. 3.

Although vowel length is also not a phonemic contrast in Mandarin, it seems that durational differentiation is easier for Mandarin learners to acquire than spectral distinctions, which is consistent with the statement in Hachenberg (2003) that Mandarin L2 learners rely mainly on durational cues to contrast German tense-lax vowels. As reported in Sec. 3.3, for the durational distance, the two speaker groups had a significant difference only for one vowel pair (/aː-a/); while, for the spectral distance, they had significant differences for six out of seven vowel pairs.

The significant difference in the durational distance for the vowel pair (/aː-a/) and the non-significant difference in the spectral tense counterpart /aː/ between Mandarin and German speakers drew our attention to an interesting phenomenon of inverse quality-duration relationship in German (Weiss, 1974). It has been claimed that for high-tense vowels, quality is more important, whereas for the low vowel pair (/aː-a/), duration serves as the primary cue. That means German native speakers rely mainly on durational and spectral distances to distinguish low and high vowel pairs, respectively; while Mandarin learners consistently use more durational features rather than spectral quality features to indicate the tense-lax contrast. As can be observed in Fig. 1, the mean vowel duration of Mandarin speakers is much longer than that of German ones for high vowels, not much longer for mid vowels, and even slightly shorter for low vowel (/aː/), because German speakers rely more on duration but less on quality when the vowel goes from high to low, but this strategy has not been observed for the Mandarin speakers.

Because German native speakers do not use spectral features to distinguish low tense vowel /aː/ from its lax counterpart, which coincides with the Mandarin speakers' performance, no significant spectral difference was found for the low tense vowel /aː/ between these two groups. Although Mandarin speakers favor durational distinctions, the distinction is not enough for the low vowel contrast. Thus, the only significant durational distance was found for the low vowel pair (/aː-a/), as reported in Sec. 3.3.

Our findings have shown the major problems that Mandarin learners of L2 German have in distinguishing German tense-lax vowels in production, and future research should investigate how we can help them to overcome these difficulties by explicit phonetic training. Moreover, although Mandarin Chinese has a simple monophthong vowel system, it has a rather complex compound vowel inventory with nine diphthongs and four triphthongs in Mandarin Chinese. It can be hypothesized that this characteristic has an influence on the L2 German acquisition. In the future, it will also be desirable to examine the production of L1 Mandarin Chinese vowels, especially the Chinese compound vowels, by native speakers together with their L2 German vowels. The present study using monosyllabic nonsense words cannot account for the influence of many factors in real speech (e.g., the number of syllables, the stress position, and orthography), which will also be considered.

This research work was partially sponsored by the China Scholarship Council and Shanghai Social Science project (Grant No. 2018BYY003). We would like to express our gratitude to the two anonymous reviewers for their constructive comments and suggestions to improve this work.