Contact geminates - test 1

The reading texts described in section 3.2.2 also formed the source from which the geminate-singleton targets were taken. The targets had not been placed deliberately in the texts. Rather, we became aware of the potential of contact geminates when labelling the acoustic data. Fifteen target words were analysed across speakers.

Targets were either a single word to derive initial, medial or final nasal singletons /n/

and /m/ from like mager/skinny, Monat/month, Guthaben/credit balance (measured segments are shown as bold) or a phrase of two words that contained nasal contact geminates /nn/ or /mm/ as inschon nach/already after. Sometimes, a target contained both nasal singletons and contact geminates: einem Monat/one month, dat. As a rule, all nasal durations within a target were measured (see Appendix, section A.1).

Furthermore, we included /nm/ as a quasi-contact geminate as in einen Mann/a man, acc.¹⁴ In contrast to a homorganic contact geminate, which consits of two succeeding homorganic phonemes, the term quasi-geminate shall describe that two akin phonemes meet, which share the manner of articulation (nasals), but their place of articulation is not identical. In the remainder of this thesis, the term “contact geminate” covers both types of contact geminates.

Since lexical frequency might influence the timing characteristics of segments (Pluymaek-ers, Ernestus, and Baayen, 2005), this factor was considered during the analysis. Lexical frequency was determined according to the frequency data provided by Wortschatz Uni-versit¨at Leipzig (http://wortschatz.uni-leipzig.de/) which comprises 35 million sentences taken form publicly accessible sources (mostly newspapers). For each word, the corpus gives two results: total frequency of the lemma and frequency class which is calculated through comparison with the most frequent German word (masc. articleder) as refer-ence. The calculation of frequency classes out of total frequency is done logarithmically and follows Zipf’s law. The law states “that the rank of a word (in terms of its fre-quency) is approximately inversely proportional to its actual frequency, and so produces a hyperbolic distribution” (Tullo and Hurford, 2003).“Der”represents frequency class 1.

Words up to frequency class 16 are regarded as common words while words belonging to frequency class 17 and above are considered as being outside the general linguistic usage (Quasthoff, 2009, p. 30). Frequency classes are dynamic, because new words enter a language while known words might be less often used due to various reasons (e.g. fash-ionable terms disappear). The highest frequency class of the corpus is 21 with extremely rare words.¹⁵

14Sometimes, quasi contact geminates were assimilated. The cited phrase einen Mann/a man, dat., for instance, was occasionally uttered as eim Mann. Quasi-geminates which were assimilated to homorganic geminates were counted as homorganic geminates.

15A word that belongs to frequency class 21, for instance, is the rarely used nounSchwippschwager which means brother of one’s brother/sister in law or husband of one’s sister in law. The word Schwippschwager has only one count in the corpus of Wortschatz Universit¨at Leipzig. The reference wordder representing frequency class 1 appears 2²¹ times more often thanSchwippschwager in the corpus.

Our target words reached from frequency classes 3 to 14 and thus belong to the area of commonly used words. Since two words are necessary to form a contact geminate, the frequency of the first and the second word were averaged. In sum, 115 nasal con-tact geminates (12 /nn/, 33 /mm/, 70 /nm/) and 166 nasal singletons (35 initial, 68 medial and 63 final) were suitable for analysis (total: 281 items). Thirty-three of the final nasals were followed by an IP boundary, which might condition lengthening of the segment. In our statistical analysis, duration was the dependent variable, condition (ini-tial singleton, medial singleton, non IP-final singleton, IP-final singleton; geminate) and lexical frequency were the independent variables and participants and items were the crossed random factors. Outliers were removed. The results showed an effect of initial strengthening, as well as word-final lengthening. Singletons in the word-medial con-dition were shorter than singletons in word-initial and -final concon-ditions (medial-initial:

beta = 0.011, lower bound = 0.0003, upper bound = 0.027, p = 0.052; medial-final:

beta = 0.028, lower bound = 0.0025, upper bound = 0.032, p = 0.072).¹⁶ Both effects approached significance. For final singletons at IP boundaries, the effect of word-final lengthening compared to medial singletons was highly significant (beta = 0.056, lower bound = 0.049, upper bound = 0.075, p<0.0001). There was no effect of lexical frequency.

Figure 3.4. Condition-induced nasal singleton durations compared to all pooled nasal contact geminates.

Initial and non IP-final singletons showed no significant difference in duration (beta = 0.017, lower bound = -0.017, upper bound = 0.023, p >0.7).

16“Beta” denotes the estimate adjustment to the intercept, “lower bound” and “upper bound” denote the lower and upper bounds of the highest posterior density interval for 95 per cent of the probability density.

IP-final singletons were significantly longer than non IP-final singletons (beta = 0.032, lower bound = 0.067, upper bound = 0.027, p < 0.0001). The duration of word-final singletons at IP boundaries did not differ from those of geminates (beta = 0.002, lower bound = -0.016, upper bound = 0.007, p >0.4).

Except for IP-final singletons, all singleton conditions differed in duration to geminates (initial: beta = 0.047, lower bound = 0.057, upper bound = 0.031, p <0.0001, medial:

beta = 0.057, lower bound = 0.066, upper bound = 0.048, p < 0.0001, non IP-final:

beta = 0.029, lower bound = 0.059, upper bound = 0.025, p < 0.0001). Durations of singletons and geminates are visualised in Figure 3.4. For lexical frequency, neither an interaction with condition nor a main effect was found.

For all speakers, the duration of the longest singleton (mostly IP-final ones) exceeded the shortest contact geminate. In order to check the potential of contact geminates as wb markers, we looked for the point at which our nasals were best dividable into singletons and geminates. We assumed an automatic classification, taking only the duration of nasals as a criterion, and found that 77 per cent of the items were correctly identifiable as singletons and contact geminates, if the separation boundary was fixed at 100 ms.

This observation shows that singletons and geminates share some overlapping area of duration. An ideal classification of all items is hardly possible despite the fact that the singleton and geminate group have clearly differing mean durations (mean singletons: 87 ms, mean contact geminates: 131 ms). However, if the separation boundary is carefully chosen, the error rate can be minimised. There is yet another method to improve the classification results, namely through the identification of singletons at IP boundaries.

Their durations are usually similar to those of contact geminates. In an automatic clas-sification process that sorts items according to segment duration, lengthened IP-final singletons would end up in the geminate group. But there is a way to identify them.

The breath group theory of Lieberman (1967) assumes that speakers tend to take a breath after the end of a syntactic unit and thus insert a silent interval. Hence, a long nasal (for our dataset, “long” means>= 100 ms) followed by a silent interval had to be a lengthened singleton. Due to phonotactic rules of German speech, a silent interval can-not follow after a nasal contact geminate. The second time slot of the contact geminate equals the first phoneme of a word. In the case of German nasals, only vowels can follow at the word onset (Yu, 1992). From an articulatory point of view, the vocal folds vibrate uninterruptedly when the sequence nasal + vowel is produced (einem Mann: nasal + vowel /m:a/). Consequently, a check of the right neighbourhood of all nasals >= 100

ms was carried out in order to find those items that were followed by a silent interval.

The procedure further improved the result for correct identification of nasal items in our dataset by 9 per cent - these are the newly covered IP-final singletons that were followed by a silent interval. The total identification score for all items in the automatic classification could thus be raised to 86 per cent.

The question of segment duration according to condition leads us to the singleton-geminate ratio, which was 1.51. If the quasi-singleton-geminates (mean: 143 ms) were taken out and only homorganic geminates (mean: 112 ms) were included in the calculation, the ratio decreased to 1.29. Quasi-geminates and homorganic geminates differed signif-icantly in their durations (beta = 0.028, lower bound = 0.014, upper bound = 0.046, p = 0.002), see Figure 3.5.

Figure 3.5. Mean durations of homorganic geminates /nn/ and /mm/ (n = 45) compared to quasi-geminate /nm/ (n = 70).

What conclusions can be drawn from the obtained results? Most importantly, nasal contact geminates proved to be useful cues to word boundaries in German speech. When two nasals framed a wb, which means that one word ended and the next one began with the same nasal, the phonemes were usually acoustically concatenated and uttered as one lengthened segment. This finding is in accord with Lea (1980) who analysed English sibilants. He explained that the exceeding of a certain threshold in the sibilant’s duration might signal a division point of two succeeding words. We found that contact geminates were significantly longer than initial, medial, and final singletons without an

IP boundary. Furthermore, we obtained segment-initial and segment-final strengthening effects. Lengthened word-final singletons at IP boundaries reached an average duration similar to contact geminates. They are identifiable when followed by a silent interval.

The quasi-geminate /nm/ was included in the study and treated like the regular ones. It was sometimes assimilated, but not as a rule. This has to be attributed to the fact that data elicitation was based on reading, which is usually characterised by a more careful articulation compared to spontaneous speech.

Our singleton-geminate ratio was below the values cited in the literature for languages with consonantal length contrast (1.61 to 2.61). While durations for /nn/ and /mm/

proposed genuine consonant lengthening resulting in a geminate, durations for /nm/

rather suggested the concatenation of two singletons.¹⁷ Astonishingly, our ratio was also below the ratio reported by Mikuteit (2007). She had found a geminate-singleton ratio of 1.82 for German alveolar stop geminates. One explanation for this divergence might be found in the material (prefixed verbs, e.g. mittesten/to test along with) and reading style of the test persons. Mikuteit’s participants had been instructed to utter isolated words with a monotone intonation, our targets were embedded in sentences. Sentence prosody might have influenced segment duration. The diverging ratios might also be due to the fact that different phoneme characteristics were observed: closure duration for stop geminates and sonority in the case of nasals.

In order to investigate the timing characteristics of singletons and contact geminates further, a second production experiment was carried out in which the number and con-dition of each singleton segment was controlled (initial, medial, non IP-final, IP-final), as well as the number of homorganic and quasi-geminates.