Schwa realisation in verbal inflection in two dialogue registers of German spontaneous speech

1 Introduction

In spontaneous speech German 1st-person singular verbs can be realised with or without final schwa. This exploratory study investigates whether the realisation of schwa is partly dependent on register differences. We investigate this by comparing task-based dialogues and free conversations, produced by the same pairs of interlocutors. As we will show, the realisation of schwa depends on many factors, such as the phonetic context or the frequency of the verb in question – but crucially it also seems to be motivated by register differences. If that is so, it posits an interesting case for register modelling.

An illustration of the variation of interest is given in Example (1), taken from the BeDiaCo corpus, which will be introduced in Section 2. Here we see that the variable habe (have.prs.1sg ‘[I] have’) can be realised with schwa (1a) and without schwa (1b).^[1]

Both variants, habe in Example (1a) and hab in Example (1b), would be represented with <habe> in standard written German, with the 1st-person singular form typically being represented with an <e>.

Since we are interested in phonetic variation, we explicitly do not assume that one of these forms is the ‘correct’ or underlying form – as we will argue in this article, different forms are distributed differently depending on the situation. We will therefore not use the (prescriptive) terms deletion (Kohler and Rodgers 2001), elision (Davidson 2006; Kohler 1990; Kohler and Rodgers 2001), or reduction (Johnson 2004; Kohler 1990; Kohler and Rodgers 2001; Lindblom 1963), which would point to an underlying phonological form with a schwa. Instead, we model verbal inflection at the morphophonetic interface as a morphological variable in Labov’s sense (Labov 1963 et passim). We will name the variable based on the written form as given in a standard grammar of German with phonetic variants of schwa realised or not realised. Accordingly, the variable ich male ‘I paint.prs.1sg’ can be expressed as ich male [ɪç maː.lə] or ich mal [ɪç maːl].

The two realisations of 1st-person singular are semantically fully equivalent (apart from avoiding homonyms with the subjunctive forms). While other phonetic variation phenomena such as the widely known English examples of ing being realised as [n] or [ŋ] or of postvocalic /r/ being realised or not (Labov 1963, 1971) are caused (at least to a certain extent) by social factors (cf. Section 1.2.4), it seems unlikely that social differences play into the distribution of the two variants with and without schwa in Example (1) above. We are not aware of any studies pointing to a social-class difference in the realisation of schwa as a verbal inflectional suffix; likewise the variants do not seem to express social meaning. As far as we know, this variation is fairly stable and does not hint at a change in progress. We investigate whether and how such “seemingly unmotivated” variation within the phonetic realisation of schwa occurs in two closely related registers. We are thus interested in the difference between realisations of 1st-person singular verbs with and without schwa. By providing an exploratory corpus-based comparison of free conversations and task-based dialogues we aim to shed light not only on how inflection is realised in both registers, but also on how the functional differences between these two situations affect verbal inflection. By focusing on the use of a small phonetic difference that mostly goes unnoticed we want to explore an interesting case for register modelling.

In this paper we will first describe the most important aspects of verbal inflection and its pronunciation in German (cf. Section 1.1). In Section 1.2, we will discuss possible factors influencing schwa realisation, with a particular focus on the factor register. Section 2 will provide a detailed description of our methodological approach and the corpus used, before presenting the results of the statistical analysis in Section 3. In Section 4, we will discuss our findings.

2 Methods

We employ a subcorpus of the Berlin Dialogue Corpus (BeDiaCo_main v.2, Belz et al. 2021), including face-to-face dialogues in two situations. All annotations marked during this study are published in BeDiaCo v.3 (Belz and Mooshammer 2023), available at the media repository of the Humboldt-Universität zu Berlin^[5] for teaching and scientific research. More details about the corpus and its annotation guidelines can be found in Belz et al. (2023).

2.4 Annotation

BeDiaCo uses a multi-layer standoff architecture with multiple segmentations. The segmentation of sound was carried out automatically using the WebMAUS service (Schiel 1999). The diplomatic transliteration (tier dipl) was normalised semi-automatically (tier norm) for the purpose of this study, based on orthographic Standard German target hypotheses. For the normalisation, all pauses, interruptions, and non-understandable as well as anonymised word tokens were excluded. Normalisation thus splits spontaneously uttered tokens on dipl such as isses ‘is it’ or hamwa ‘we have’ into two word tokens on norm which are ist es and haben wir, respectively (cf. Figure 1).

Figure 1:

Example of the annotation layers in BeDiaCo_main. The layer names are on the right. In this case, we expect schwa (“e”) on the layer flexexp, which is not realised (“f” on flexreal). The syllable of the inflected verb in which schwa is not realised is cliticised (“c” on stress), and the following context is therefore re-categorised as unstressed.

Further, part-of-speech tags (POS) based on STTS 2.0 (Westpfahl 2014; Westpfahl et al. 2017) and lemmas were created fully automatically with the help of TreeTagger (Schmid 1994) and inserted as additional annotation layers lem and pos. Morphological information, such as person, number, tense, and mood, was added automatically with the help of DEMorphy (Altinok 2018) and CISTEM (Weissweiler and Fraser 2018) to all 6,045 verbs. A Python program was used to assemble these layers in a single step (Lange 2021). Subsequently, norm, pos, and gram were corrected manually and extended by layers that list the inflectional suffixes as given in the standard German grammar (flexexp) as well as the actual realisation of these (flexreal), and detailed phonetic information on the realisation (flexmau, based on the WebMAUS segmentation; cf. Figure 1). All annotations were published in Version 3 of BeDiaCo.

The realisation of schwa in verbal inflectional suffixes was assigned to two categorical tags during the manual annotation process: full and no realisation of schwa. The inflectional suffix <-e> was annotated as a full realisation if the acoustic signal showed clear periodic stretches for the schwa. If the following word token began with a vowel, the following criteria were used to set boundaries to the next segments: the movement of the first and second formants, and/or glottalisation or glottal closure. The tag no realisation was used if there was no schwa visible in the sonagram or in the audio signal in the form of periodicity and/or clear change of F1/F2 structure to the following segments. The annotation criteria for the past tense suffix <-te> were the same for the schwa, while another annotation tag, partial realisation, was used if only a part of the schwa ending was realised (<-te> [tə] to <-t> [t]). Additionally, the POS tags for haben ‘to have’ (starting with either VV for lexical verbs or VA for auxiliaries) were corrected manually, since the tagger does not distinguish between the use as a lexical verb or auxiliary verb, with auxiliary being assumed in most cases.

The intonation phrase tier was annotated manually. The boundary of an intonation phrase was marked obligatorily if the phrase ended with a breath pause or a boundary tone. Additionally, the phrasal boundary could be signalled by segmental lengthening, laryngealisation, tonal movement, or a silent pause longer than 50 ms (Belz 2021: 81–90, in detail Belz et al. 2023). Finally, the factor stress of the following syllable was manually annotated for all verbs on an additional tier based on the auditory impression.

3 Results

Overall, schwa is realised in 23.7 % and not realised in 76.3 % of all 855 instances. These numbers are not equally distributed across registers (cf. Table 3). The free dialogues show a significantly higher number of schwa realisations than the task-based Diapix dialogues (χ² = 15.8, df = 1, p < 0.001).

Table 4 shows the number of schwa realisations per situation for each categorical predictor. We see that the task-based Diapix dialogues behave differently with respect to tense. In detail, only 3 of 182 verb forms in the Diapixes are realised in preterite. Other factors also show very low instances in some categories; for example, there are no instances of verb forms with schwa used in auxiliaries in the Diapixes. It seems that present tense, lexical verbs, and verbs next to ich are indicative of the speech used to solve a Diapix task. This is in line with our pre-assumption that there will rarely be any preterite cases in this situation. As the three instances in the preterite category of tense do not warrant a statistical analysis, we exclude tense from the models.^[7]

In order to test whether the significant difference in schwa realisation per register (higher occurrences in free than in Diapix conversations) will be persistent when including segmental, suprasegmental, morphosyntactic, and frequency factors known to affect schwa realisation, we build two generalised linear mixed-effects models with schwa realisation as the binary response variable. The first model contains all IP instances, while the second model only contains IP-medial instances. This is motivated by the fact that the IP position is confounded with following context, as pauses following a verb form with potential schwa often coincide with IP-final position (54 out of 88). The same applies to stress. The first model (theoretical conditional variation explained: R _c ² = 0.39) includes all verb forms. Paradigmatic compensation, global articulation rate, local articulation rate, and part-of-speech do not improve the model significantly. No meaningful interactions between factors were identified.

Model 1: realisation ∼ IP position + adjacent ich + situation + lemma frequency + (1|lemma) + (1|speaker)

The second model (theoretical conditional variation explained: R _c ² = 0.43) only includes verb forms which are not in the final IP position. Again, paradigmatic compensation, global articulation rate, local articulation rate, and part-of-speech do not improve the model significantly. Tense is again excluded due to data scarcity. The interaction between stress and following context was included, as it improved the model significantly.

Model 2: realisation ∼ adjacent ich + situation + lemma frequency + stress of the following syllable * following context + (1|lemma) + (1|speaker)

In order to determine how much variance is explained by the factor situation, we compared R _c ² of both models to R _c ² calculated for models without situation. Situation thus explains an additional 2.8 % of the variance for Model 1 and 3 % for Model 2, respectively.

Table 5 summarises the predictions of both models. The models are comparable with respect to their changes in predictors. As with schwa is the reference level represented with 0 in the model, the negative slope indicates a movement towards schwa realisation.

In Model 1, non-final IP position leads to significantly higher amounts of unrealised schwa, meaning that schwa is significantly more likely to occur at IP boundaries than within intonation phrases (see also Figure 2a). Turning to Model 2, schwa is realised significantly more often in free than in task-based situations (see Figure 2b). If the verb form is preceded or followed by the pronoun ‘I’, schwa is significantly less often realised than if the verb form does not occur together with ‘I’ (Figure 2c). It should be noted that 305 out of 676 cases involve a following ‘I’, as in hab ich ‘have I’. Frequent lemmas lead to significantly fewer realisations of schwa, meaning that infrequent verbs tend to be realised with schwa (Figure 2d). Finally, if the next syllable starts with an unstressed vowel, schwa frequency is reduced significantly (Figure 2e).

Figure 2:

Predicted marginal effects proportions of verb forms realised without schwa (95 % confidence intervals) for (a) IP position in Model 1. Plots (b)–(e) show the effects of Model 2 for situation, adjacent ich, lemma frequency (back-transformed), and following context in interaction with stress.

Neither global nor local articulation rates improve the model of schwa realisation as presented above. Although interlocutors tend to speak faster (globally) in the free dialogues (5.5 syllables/second) than in the Diapix tasks (4.9 syllables/second), this is irrelevant to the realisation of schwa. Evidence is given in Figure 3, showing no effect of the local articulation rate per verb form on schwa realisation in both situations. This means that interlocutors do not speak faster within the same intonation phrase in which they realise verb forms without schwa compared to realising verb forms with schwa.

Figure 3:

Local articulation rate per verb form with schwa (grey) and without schwa (white) in Diapix tasks and free dialogues.

4 Discussion

In summary, our results show that the variant without schwa is the most frequent realisation in verbal inflectional endings for the 1st-person singular (76.3 % of all cases with potential schwa, n = 652). Final schwas are promoted by occurrence before IP boundaries, in the preterite, and in less frequent verbs. Furthermore, verbs that are not preceded or followed by ‘ich’ are also produced more frequently with schwa. Paradigmatic compensation, articulation rate, and verb type (auxiliary or lexical verb) did not affect the realisation of schwa. Concerning our research question, i.e. whether register is also a factor affecting the frequency of inflectional schwa realisations, our results indicate that schwa realisations are significantly more likely in free conversations as opposed to task-based dialogues. In this section, we will first discuss the effect of register and then the other factors, also in relation to register.

In this study, two closely related registers of unscripted spontaneous speech are compared in regard to schwa realisation, task-based versus free conversation. The ‘mode’ and the ‘tenor of discourse’ (Biber and Conrad 2019; Halliday 1978) are kept identical for the two registers. Since both interlocutors are students about the same age, the speakers in this study have a similar socio-economic background and therefore effects of social relationship (see, e.g. Eckert and Labov 2017; Labov 1971) are minimised in our corpus.

The variants investigated here, the verbal inflection for 1st-person singular with and without schwa, are semantically equivalent and – at least to our knowledge – they do not carry any social meaning. Therefore, we can exclude these factors as possible triggers for the observed difference in schwa realisations between free and task-based conversations. These two similar registers differ in their content and potential communicative purpose (‘field of discourse’) (Biber and Conrad 2019; Halliday 1978).

In the introduction, we speculated that due to higher processing costs speakers speak slower in the Diapix task compared to the free conversation. Indeed, we found the global and the local articulation rate to be faster in the free dialogues than in the Diapix dialogues, reflecting the higher cognitive demands needed during solving the task of finding differing pictures, and corroborating Watson et al. (2020). As stated by, e.g. Lindblom (1990) and Kohler (1990), fast speech leads to shorter word and segment duration and to more frequent reduction phenomena such as assimilation, segment elisions, and vowel reduction in read speech (cf. Browman and Goldstein 1990; Davidson 2006; Hoole and Mooshammer 2002), as well as in spontaneous speech (cf. Bell et al. 2003; Ernestus 2014; Fosler-Lussier and Morgan 1999; Hanique et al. 2010; Raymond et al. 2006). However, increased time pressure does not automatically lead to more frequent ‘reductions’, see e.g. Van Son and Pols (1990, 1992, Kienast and Sendlmeier (2000), and Koreman (2006). Ernestus (2014) and Ernestus et al. (2015) argue that phonetic reduction phenomena are under the control of the speaker since – as they found – ‘reductions’ can also depend on the socio-economic status of the speaker. Our results give further evidence for this non-mechanistic view: With free conversations exhibiting a faster articulation rate while featuring more realisations with schwa compared to the Diapix task, our results indicate that a faster articulation rate does not necessarily lead to fewer realisations with schwa, contrary to the findings of, for example, Hanique et al. (2010) and Ernestus et al. (2015), as well as the assumptions of Kohler (1990).

Other factors also affected the frequency of schwa realisations, such as IP position, word frequency, stress of the following syllable, the following phonological context, and adjacent ich, but these factors cannot be discussed independently of the investigated registers. For example, as predicted from the literature (Belz et al. 2022; Niebuhr et al. 2013; Piroth and Janker 2004), we also found more frequent schwa realisations in IP final position, indicating phrase-final strengthening (Byrd and Krivokapić 2021; Cho 2011). However, the position of the finite verb within prosodic phrases is also distributed in a register-specific way: In task-based dialogues, only eight instances of the verb forms investigated in this study were in phrase-final position, compared to 80 in free conversations (see Table 4). A similarly uneven distribution was observed for adjacent ich, part-of-speech and tense: Only very few instances of verbs without ich, auxiliary verbs and verbs in preterite occur in the Diapix tasks. Even though these skewed distributions hampered the statistical analyses, the distribution of these factors is strongly influenced by the register, suggesting they act as linguistic features constituting a register (Lüdeling et al. 2022). We find a clear preference for non-final IP position, present tense, lexical verbs, and occurrences with an adjacent ich in the Diapix task: Participants seem to produce similar sentence structures, with the finite (lexical) verb in the present tense in second position within the sentence, and seem to vary little. Since the Diapix task is firmly situated in the present, the usage of mostly present tense and few auxiliaries (with auxiliaries usually being used in the periphrastic past tense; cf. Section 1.1) is perfectly plausible. The time pressure in the Diapix task might lead to less complex^[8] and less diverse sentence structures in order to ensure the interlocutor’s immediate understanding. Hence, these features follow the specific demands of the communicative situation. In contrast, free conversation exhibits more features that tend to promote realisations with schwa. With no task-specific time pressure and a generally more leisurely conversation flow, speakers are more flexible and vary more in free conversations.

The factor following context is interesting with respect to the pronoun ich ‘I’ frequently following the verb, in which case the verb form tends to be realised without schwa. As the personal pronoun ich in spontaneous speech is mostly unaccented and starts with a vowel (only 10 % of all ich tokens are produced with a glottal stop before the vowel in Wesener 1999), the realisation without schwa could either be due to a hiatus avoidance strategy, or a resyllabification of the last segment of the verb stem. As we find hab ich ‘have I’ very frequently in the data, it may be hypothesised that these words form a collocation or construction. In the most frequent sequences (i.e. in the context of an adjacent ich ‘I’), verb-final schwa is less likely to be realised than in less frequent sequences without ich – indeed, we find very few realisations with schwa in these frequent sequences, supporting Hanique et al. (2010) and Kohler (1990). As suggested by Hall (1999), the verb stem followed by ich forms a single prosodic word, evidenced by the resyllabification in [hab̥ɪç]. Our results indicate that verb-final schwas are more frequently realised in less predictable contexts, suggesting that frequent sequences form strong links that hinder schwa realisations. This effect could be due to both cliticisation or higher predictability in more frequently occurring contexts. Cliticisation, however, would only explain cases with following ich that are realised without schwa (94.8 % of 308) and not cases with preceding ich (76.5 % of 412). It is possible that cliticisation effects and predictability effects interact, with more clitics in more predictable contexts leading to fewer realisations with schwa. Cliticisation might also explain the significant interaction between stress and following vowels with less frequent realisations before an unstressed vowel, most of which are the vowel in ich (292 out of 345, 95.9 % without schwa). In contrast, in less frequent sequences (cases without an adjacent ich) speakers seem to be more flexible and vary more in their schwa realisations. This tendency, however, holds true for free conversations only. In the Diapix task, we find very few cases without adjacent ich (n = 9, six cases without schwa), which hardly allows conclusions to be drawn. Generally, one of the shortcomings of our study is that some cases occur very rarely (e.g. three cases of the preterite in the Diapix task), rendering our analysis vulnerable to individual idiosyncrasies.

Further, paradigmatic compensation does not significantly affect the distribution of schwa realisation in our study. Paradigmatic compensation predicts more frequent schwa realisations for stems ending in nasals or obstruents, e.g. hab-, glaub- and find-, compared to stems with final vowels, diphthongs or liquids, e.g. geh-, bau- (Eisenberg 2013; Raffelsiefen 1995). Since schwas are rarely realised in the most frequent verb haben, we can assume that the paradigmatic compensation does not play a role in unscripted spontaneous speech registers. The examples above also give evidence that avoidance of final devoicing is irrelevant for the speakers investigated in our study. Similarly, the effect of rhythmicity, i.e. a preference for alternating stress patterns (see Kentner 2018), does not affect schwa realisation per se. However, inflected verbs have a significantly higher probability of schwa realisation when the following syllable starts with a vowel and is stressed rather than unstressed. Apart from this case, speakers in our study are not sensitive to stress clash and do not strive for a repair by deleting or inserting a schwa. Up to now, evidence for this factor is restricted to written texts and read speech for German (Fleischer et al. 2018; Kentner 2018).

5 Conclusions

In general, word-final schwa is rarely realised in verbal inflectional suffixes, a phenomenon which we did not expect to be so pronounced. This indicates that realisations without schwa are more common in German spontaneous speech than realisations with schwa – implying the need to reconsider standards and norms in spontaneous speech, especially in second language acquisition and automatic speech processing: Word-final schwa in verbal inflectional suffixes in present tense should be taught as at least optional if not omissible. While suprasegmental, morphosyntactic, and frequency factors influence schwa realisation, register, too, proves to be a crucial factor that needs to be considered in modelling. As argued in Lüdeling et al. (2022), even very fine-grained, functional differences in register affect the distribution of register features, which in turn influence the realisation of word-final schwa. Thus, when analysing seemingly free variation of (morpho)phonetic phenomena, i.e. variation that is not motivated by semantic, pragmatic or social-meaning differences, the factor situation, through which at least grammatical factors (e.g. tense) are mediated, must also be taken into account. It is essential to distinguish the influence of conventionalised text genres on schwa realisation from the situation or register, particularly when more regular speech and a constant rhythm are required, such as in poetry. Furthermore, the relationship between articulation rate and ‘reduction’ phenomena also seems to vary with register. It will be a fruitful future endeavour to investigate this relationship for different registers varying with respect to formality, social distance, or addressee.