1.2.1. The Norwegian merger

Previous research on the Norwegian merger of /ʃ/ and /ç/ has used varying methods of identifying merged and distinct speakers, most of which have relied on perception. Although these studies provide some useful insights into the behavior of potential merged and distinct speakers, they have certain methodological weaknesses that potentially affect their conclusions.

Papazian (1994) conducted a study of the distribution of the merger among 11, 15 and 18 year olds in Oslo using a questionnaire and a reading test. The questionnaire asked speakers to state whether they pronounce eight minimal pairs containing /ʃ/ and /ç/ the same or differently. In the reading test, a subset of the speakers who filled out the questionnaire were asked to read aloud sentences containing these minimal pairs. Papazian then noted whether he believed that the speaker had produced /ʃ/ or /ç/ for each word. This reading test was intended to be a control for whether or not the answers given in the questionnaire accurately represented the speakers’ productions.

This method of finding merged and distinct speakers is vulnerable to possible biases of the speakers and the researcher. It is widely recognized that when conducting linguistic research, asking people how they speak does not necessarily give an accurate representation of how they actually speak (Labov, 1972, p. 213). This is particularly the case when the object of study is sociolinguistically disfavored, as is the case with the Norwegian merger (Torp, 1999, p. 344). Furthermore, researchers might be subject to confirmation bias (Nickerson, 1998), leading to selectivity in the interpretation and use of evidence to support a hypothesis that the researcher already believes to be true.

Similarly, Dalbakken (1996) relied on her own perception to identify merged and distinct speakers in her study of spontaneous and read speech among teenage speakers from Trondheim. These data were subsequently used by van Dommelen (2003) in an acoustic study of the merger. Van Dommelen made some slight modifications to this classification based on his own judgment of the fricatives, and his analysis therefore relied on his and Dalbakken’s perception. Relying on the perception of two listeners is vulnerable when classifying the productions of /ʃ/ and /ç/. Specifically, in those cases where the two researchers disagreed, the reader cannot know which of them is more accurate. This point can be anecdotally illustrated by the fact that in our own research on the merger of /ʃ/ and /ç/, the two native speaker authors of this paper have disagreed on the classification of many of the tokens.

In a later acoustic analysis of the merger of /ʃ/ and /ç/, van Dommelen (2019) attempted to improve on this method. In this study, he used a speech corpus including speakers of various Norwegian dialects producing both read sentences and spontaneous speech. To identify merged and distinct speakers, van Dommelen used the perception of multiple listeners to categorize the speakers’ productions. Specifically, van Dommelen asked seven native Norwegian speakers to rate whether the productions of /ç/ were realized as [ç] or [ʃ]. Importantly, however, van Dommelen only used this method when classifying the productions of /ç/, not the productions of /ʃ/. He states that all productions of /ʃ/ were realized as [ʃ], but it appears that this conclusion was reached based only on his own perception. As stated in Section 1.1, the outcome of the merger has not been sufficiently documented, and it is possible that some speakers merge the fricatives in the direction of /ç/. Because van Dommelen used his own perception alone when classifying the productions of /ʃ/, such speakers might have been overlooked.

Finally, similarly to van Dommelen (2019), Andrésen (1980) used a perception test to investigate the Norwegian merger of /ʃ/ and /ç/. Andrésen recorded three 13-year-old speakers from Bergen, producing two sets of minimal pairs containing /ʃ/ and /ç/. These productions were subsequently categorized by the three speakers and two researchers trained in phonetics through a perceptual identification task. That is, for each realization, the listeners were asked to report whether they heard /ʃ/ or /ç/. Importantly, the productions of both /ʃ/ and /ç/ were rated. Andrésen did not have an expressed goal of determining whether the three speakers were merged or distinct speakers, but the results of the perception study reveal how accurately listeners could classify speakers’ productions of /ʃ/ and /ç/. Speakers whose productions were mostly identified correctly probably had a distinct pronunciation, whereas speakers whose productions were often misperceived most likely had a merged pronunciation. This study thus demonstrates a possible approach to identifying merged and distinct speakers using a perception study.

1.2.2. Other mergers

In attempting to explore whether acoustic measurements and perception studies allow us to identify merged and distinct speakers in Norwegian, it is relevant to assess the methods that have been used to find two such groups in the study of other ongoing mergers. The merger of /ʃ/ and /ç/ in German and the merger of alveolars and retroflex sibilants in Taiwan Mandarin serve as relevant examples from other languages (for other studies investigating the degree of merging among speakers in a situation of an ongoing merger see, e.g., Cheng et al. [2022] on Hong Kong Cantonese, Hay et al. [2006] on New Zealand English, and Irons [2007] on Kentucky English).

Jannedy and Weirich (2017) investigated the production and perception of /ʃ/ and /ç/ in certain varieties of German, where, as in Norwegian, the two fricatives are merging. They recorded speakers from three areas: Kreuzberg (a neighborhood in Berlin), Berlin (excluding Kreuzberg), and Kiel. The merger of /ʃ/ and /ç/ is typically associated with the multiethnolect variety of German spoken in Kreuzberg, while speakers from other parts of Berlin and Kiel are not expected to merge the fricatives to a notable extent (Jannedy & Weirich). The speakers were recorded producing minimal pairs containing the fricatives /ʃ/ and /ç/, and in a perception task, 12 listeners were then given the task of identifying the recorded words. The accuracy with which listeners correctly identified the speakers’ productions was calculated for each speaker group, and the results for the Kreuzberg speakers were compared with the results for the Berlin and Kiel speakers.

Jannedy and Weirich (2017) thus based their initial categorisation of speakers on dialectal and sociolectal background. Speakers were grouped as Kreuzberg speakers, Berlin speakers, or Kiel speakers, and the perceptual identification task was conducted to assess the accuracy of identification within these groups. Individual speakers’ degree of merging was not investigated, and so there might be speakers in the Kreuzberg group who did not merge and speakers in the Berlin and Kiel groups who did. This approach is useful if the purpose of the investigation is to gain an overview of trends within different groups of speakers. However, if the purpose of the study is to identify the degree of merging in individual speakers, one would have to investigate the accuracy scores for each speaker, similar to the method used by Andrésen (1980) in his small-scale study of the Norwegian merger.

Moving on to another merger, Lee-Kim and Chou (2022) investigated the so-called sibilant merger in Taiwan Mandarin. Whereas Standard Mandarin has a contrast between the alveolars /s ts tsʰ/ and the retroflexes /ʂ tʂ tʂʰ/, Taiwan Mandarin speakers often merge this contrast, resulting in a deretroflexion of the retroflex category (Kubler, 1985). Lee-Kim and Chou recorded speakers of Taiwan Mandarin producing four words containing /s ʂ tsʰ tʂʰ/, and the recordings were then analyzed acoustically. The mean of the spectral energy distribution (center of gravity) was measured for each fricative. For each speaker, the difference in spectral mean between the alveolars and the retroflexes was calculated. The resulting value was referred to as spectral distance, and Lee-Kim and Chou used this value to identify merged and distinct speakers.

Using spectral distance as a measure of speakers’ degree of merging builds on previous research on the acoustics of voiceless fricatives in this variety. The spectral mean, the acoustic parameter examined by Lee-Kim and Chou (2022), has been found to be an important acoustic parameter to the distinction between alveolars and retroflexes in Taiwan Mandarin (Lin & Wu, 2023). This parameter has also been found to pattern with listeners’ perception of the sibilants /s ʂ/ (and /ɕ/) in an identification task with Taiwan Mandarin speakers and listeners (Chiu et al., 2020).

The results of the production study indicate that speakers formed a continuum from speakers with a spectral distance around 0—indicating a merged pronunciation—to speakers with a spectral distance of 4000 Hz—indicating a distinct pronunciation. To divide the speakers on this continuum into two groups, Lee-Kim and Chou (2022) investigated the spectral energy distribution of alveolars and retroflexes for each speaker. They then investigated the differences between these distributions using a Bayes factor test. Speakers who had no significant difference between the two categories were classified as merged speakers, and speakers who had a significant difference were classified as distinct speakers.

This approach thus relies on acoustic measurements rather than perception in the identification of merged and distinct speakers. An important factor in this analysis is that previous research has investigated which acoustic parameter is most relevant to the distinction at hand. The speakers in Lee-Kim and Chou’s study were ranked in a certain order according to spectral distance, from speakers with smaller values (the merged speakers) to speakers with larger values (the distinct speakers). However, it is not a given that this order would be the same if one were to investigate the differences between alveolars and retroflexes based on other acoustic parameters. In attempting to identify merged and distinct speakers, then, it is an advantage to know which acoustic parameter is most important in signaling the relevant contrast in the language or variety under investigation. If this is not known, it is necessary to measure a range of acoustic parameters and investigate their importance to the contrast at hand.

However, even if it is possible to obtain a ranking of participants based on the size of their acoustic distance for a given parameter, it is not necessarily the case that one can divide these speakers into groups in a non-arbitrary way. In Lee-Kim and Chou’s (2022) study, speakers’ spectral distances formed a continuum, and there was no obvious cutoff point for inclusion in the merged and distinct speaker groups. A boundary between speakers was obtained using statistics, but because there was no evidence for a clear separation of speakers in the data, such a boundary is not necessarily informative. If speakers form a continuum rather than separate groups, then, it is possible to categorize speakers, but not without making arbitrary decisions.

2.1.1. Participants

Forty-two native speakers of Norwegian participated in the experiment (32 female, 10 male; aged 18–52, mean age 24). Because the realization of /ʃ/ and /ç/ varies between different Norwegian dialects, /ʃ/ being produced as [sj] in certain dialects and /ç/ being produced as [tʃ] in certain dialects (Torp, 1999), participants’ native dialect was restricted to East Norwegian. In this dialect, both phonemes are invariably produced as fricatives. Participants were classified as speakers of East Norwegian if their home municipality was within a 120 kilometer radius from the city center of Oslo. Furthermore, only participants whose only native language is Norwegian, and whose parents also have Norwegian as a native language, were included. Participants were first-year students of Scandinavian studies completing an introductory course in Norwegian grammar, and they received course credits for participation.

2.1.2. Materials

Participants were asked to read out loud a set of nonce words of a CV structure. To conceal the purpose of the experiment, the nonce words contained not only /ʃ/ and /ç/ as C, but all possible consonants of Norwegian. As V, the nonce words contained either of the vowels /i ɑ u/. This resulted in 51 nonce words (17 consonants × 3 vowels), and each of these was read twice. Each participant thus read 102 words in total.

The purpose of using nonce words rather than real words was to access speakers’ abstract categories for the phonemes /ʃ/ and /ç/, rather than their word-level representations of real words. Furthermore, using nonce words allows us to control for lexical frequency to a greater extent, a factor which has been shown to affect the rate and diffusion of sound changes (Todd et al., 2019). Some of the nonce words happened to be real words of Norwegian, such as /sɑ/ and /ʃi/. However, the nonce words were presented in such a way that there was no indication that the words carried any meaning. They were also not necessarily spelled in the same way as the corresponding Norwegian word. For example, the Norwegian word /ʃi/ ‘ski’ is spelled ‹ski›, while the nonce word /ʃi/ was spelled ‹sji› (see Section 2.1.3). These factors are likely to minimize the possible effect of nonce words corresponding to real words.¹ The usefulness of using nonce words in the investigation of mergers has been demonstrated by Hay et al. (2013).

2.1.3. Recording

The participants were recorded in the Sociocognitive Lab at the University of Oslo. The recordings were made using a head-mounted AKG MicroMic C544 L condenser microphone and a Focusrite Scarlett 2i2 3rd generation audio interface. The recordings were made in Audacity version 3.4.2 with a sampling rate of 41 000 Hz. Each participant was seated opposite the researcher and informed about the task they were to complete. Participants were then placed in front of a laptop where they were shown a PowerPoint presentation. Each slide in the presentation contained one nonce word, presented in a randomized order, and the participants were asked to read the word out loud. The presentation ran automatically once the participant clicked start, and the 51 nonce words were displayed with a two-second interstimulus interval. Participants were then offered to take a break before the 51 nonce words were repeated in the same manner, in a new randomized order. Each participant was presented with a different order.

The orthographic forms ‹sj› and ‹kj› were used to prompt participants to produce /ʃ/ and /ç/. The two fricatives can be spelled in different ways in Norwegian, and no letter can signal /ʃ/ or /ç/ alone. The forms ‹sj› and ‹kj› are the most common representations of these phonemes, and they are often referred to as the sj-sound and the kj-sound among the general public. What’s more, ‹sj› and ‹kj› unambiguously signal /ʃ/ and /ç/, respectively, in Norwegian. For these reasons, this orthography was chosen.

2.2.1. Segmentation

The fricatives /ʃ/ and /ç/ were identified in the recorded sound files and subsequently segmented and annotated by the first author using Praat (Boersma & Weenink, 2022). Because it was expected that some speakers would have the merger, it was necessary to consult the PowerPoint presentation presented to each participant to determine which fricative the participant had been instructed to produce. The annotations reflected the intended fricative rather than the perceptual impression. Both the waveform and the spectrogram were consulted when identifying the boundaries of the fricatives. The onset of each fricative was marked at the onset of high-frequency energy as indicated in the spectrogram, and at the onset of rapid zero-crossings in the waveform. Similarly, the end of the fricative was marked at the offset of high-frequency energy and the zero-crossing preceding either the onset of periodicity associated with the following vowel or a period of aspiration. Norwegian fricatives are not phonologically aspirated, but in the recorded materials, many of the fricatives appear to be phonetically aspirated, possibly as a result of hyperarticulation. This aspiration was identified as a period of low-intensity energy following the offset of the fricative and preceding the onset of periodicity of the vowel. Aspiration was not segmented as part of the fricative.

In total, the experiment resulted in 504 tokens (42 speakers × 2 fricatives × 3 vowel contexts × 2 repetitions). However, three speakers were excluded because they reported to have a speech impairment affecting the production of sibilants. Furthermore, some speakers were unsure of how to interpret the ‹sj› and ‹kj› prompts, and they therefore produced these sequences as [sj] and [kj] rather than [ʃ] and [ç] in some cases. One of these participants was excluded because all of their /ç/ tokens were produced as [kj]. The other participants who occasionally produced [sj] and [kj] were included in the analysis, although these specific tokens were excluded. The exact number of productions of each fricative in each vowel context thus varied slightly between participants. The resulting number of fricatives submitted for analysis, excluding the [sj] and [kj] tokens, was 449, produced by 38 speakers.

In addition to the fricatives, the vowel following the fricative was segmented and annotated. The onset of the vowel was identified as the onset of periodicity following the preceding fricative, or in those cases where the fricative was phonetically aspirated, following the aspiration. The end of the vowel was identified as the offset of periodicity at the end of the word. The purpose of segmenting the vowels was to conduct analyses of the second formant (F2) in the transition between the fricatives and the following vowels.

2.2.2. Acoustic analysis

Prior research on which acoustic parameters contribute most to the distinction between /ʃ/ and /ç/ in Norwegian is limited. Only van Dommelen (2003, 2019) has investigated the acoustics of these fricatives, measuring duration, relative root-mean square (rms) amplitude, and spectral characteristics such as the spectral peak and center of gravity. However, as argued in Section 1.2.1, van Dommelen’s analyses in large part rely on his own perception in categorizing speakers’ productions of /ʃ/ and /ç/, possibly affecting the results. Previous research on the distinction between /ʃ/ and /ç/ in German has investigated spectral moments and discrete cosine transformation (DCT) coefficients and found that DCT coefficients appear to more accurately capture the acoustic differences between the fricatives (Jannedy & Weirich, 2017). Research on Luxembourgish, however, measured the center of gravity and spectral peak of /ʃ/ and /ç/, finding that these parameters adequately differentiate between the fricatives (Conrad, 2023).

Looking at previous research examining voiceless fricatives in general, most studies have measured all or a subset of the following acoustic parameters: spectral moments, spectral peak, duration, F2 transitions, and intensity or rms amplitude (e.g., Forrest et al. [1988] and Jongman et al. [2000] on English, Nirgianaki [2014] on Greek, and Wikse Barrow et al. [2022] on Swedish). Because of the scarcity of research on Norwegian, and for comparability with studies of voiceless fricatives in a range of languages, we chose to focus on these acoustic parameters in the current study.

Starting with the first four spectral moments, then, these measure a spectrum’s energy distribution, where the spectrum is treated as a probability density distribution (Forrest et al., 1988). The measurements consist of 1) the spectral mean, often referred to as the center of gravity, representing the mean of the energy distribution, 2) the spectral standard deviation, referring to the spread of the energy with regard to the mean, 3) the skewness of the distribution, indicating whether the distribution is asymmetric towards higher or lower frequencies, and 4) the kurtosis of the distribution, referring to whether the distribution is more or less peaky than a normal distribution.

In addition to spectral moments, we measured the fricatives’ duration and intensity, and the F2 at the point of transition between the fricative and the vowel (Jongman et al., 2000). Furthermore, because the fricatives showed signs of phonetic aspiration, the duration of aspiration and its distribution across fricatives were analyzed.²

Spectral moments were analyzed in Praat using a script developed by DiCanio (2013). The script computed discrete Fourier transforms (DFTs) from a set of windows placed across the duration of the fricative. The windows were then time-averaged (Shadle, 2012), meaning that the DFTs were averaged for each token before measuring and collecting the spectral moments. This way, the measurements were not as sensitive to fluctuations in the signal as they would be if they were based on one window. In our analysis, we extracted 6 Hann windows of 15 ms evenly spaced across the middle 80% of the fricative. The script used a low-pass filter of 300 Hz. In addition to the spectral moments, the script measured the average intensity across the extracted windows and the total duration of the fricative.

Aspiration was measured using a modified version of the Praat script developed for the acoustic analysis in Wikse Barrow et al. (2022). This script measured the duration of the interval between the offset of the fricative and the onset of the following vowel. In those cases where the fricative was not phonetically aspirated, then, aspiration was measured to be zero. As for F2 transitions, these were measured using the script developed by García (2017). Measurements were collected from a 25 ms window placed at the onset of the vowel.

Two of the measured acoustic parameters, aspiration and kurtosis, showed positively skewed distributions and were therefore log-transformed (see Winter [2020, pp. 90–91] for a discussion of log-transformation of skewed data). The kurtosis variable contained negative values, and consequently a constant was added to each measurement prior to the log-transformation. The log-transformed measurements for aspiration and kurtosis were used in all subsequent analyses in the study. Moreover, in the acoustic analysis, all parameters were standardized for speaker and vowel using z-scores to avoid speaker-specific and vowel-specific idiosyncrasies influencing the analysis and to allow for comparison across the different acoustic parameters. Z-scores were calculated by subtracting the mean and dividing by the standard deviation within each grouping, with the result that all variables have a mean of zero and a standard deviation of 1.

2.2.3. Statistical analysis

The contribution of each acoustic parameter to the distinction between /ʃ/ and /ç/ was estimated by modeling the effect of fricative on the different parameters and subsequently ranking the estimates. Parameter estimation and inference was performed within the Bayesian framework. In Bayesian models, we approximate a posterior distribution, that is, a distribution of plausible parameter values given the data, the model and any prior assumptions about the estimated parameter values. Because the approximation of a posterior distribution is computationally costly, we use algorithms to sample values from the posterior distribution. The practical consequence of this choice is that we do not calculate a single point estimate for a parameter, but, in our case, we draw a sample of 8,000 plausible values for that parameter and quantify our uncertainty about it by summarizing the distribution of values (see Vasishth et al., 2018; Franke & Roettger, 2019, for tutorials using phonetic examples).

To examine which acoustic parameters speakers use to signal the contrast between /ʃ/ and /ç/, we fitted a set of Bayesian linear mixed models to the production data. The models were computed in R version 4.4.1 (R Core Team, 2021) using the package brms version 2.21.0 (Bürkner, 2017). The brms package allows the user to compute models with Stan, making use of Markov Chain Monte Carlo (MCMC) sampling (Stan Development Team, 2024). The R script used for the analysis, including the computation of models, standardization of variables, and plotting, was adapted from Roessig et al.’s (2022) publicly available script.

Separate models were computed for each of the eight acoustic parameters, following the analysis conducted in Roessig et al. (2022). The measurements of the acoustic parameters were standardized using z-scores before fitting the individual mixed-effects models. As stated in Section 2.2.2, z-scoring has the effect that all parameters have a mean of 0 and a standard deviation of 1, making it possible to compare the beta weights from each model. Each model contained the relevant acoustic parameter as the response variable, while fricative (/ʃ, ç/), the following vowel (/i, ɑ, u/), and the interaction between them were entered as fixed effects. Because the nonce words were made up of only the fricative under investigation and the following vowel, the variable item was not entered as a random effect in the models. As random effects, the models included by speaker varying intercepts and varying slopes for fricative by speaker. The model structure was thus as follows:

$\mathit{\text{acoustic}} \mathit{\text{parameter}}\sim \mathit{\text{fricative}}*\mathit{\text{vowel}}+\left(1+\mathit{\text{fricative}} | \mathit{\text{speaker}}\right)$

The approach of fitting separate models for each parameter rather than entering all the parameters as predictors in one model was chosen for two reasons: First, the different acoustic measures are correlated to differing degrees, leading to collinearity. The correlation matrix in Figure 1 below illustrates the correlation between the eight acoustic parameters measured in the production study. The strongest correlation is found between spectral standard deviation and kurtosis, showing a correlation of r = –0.74, but several other parameters show correlations up to r = 0.63. As illustrated in Figure 3(a) in Schertz and Clare (2020, p. 5) (adapted from Clayards [2018]), correlated acoustic cues can play independent roles in signaling a contrast between two categories. However, correlations between parameters pose difficulties for statistical modeling. Specifically, collinearity between two or more predictors in a mixed model renders the model results uninterpretable. While there are suggested “remedies” for collinearity, statistically sound approaches such as random forests and supervised component regression come with their own limitations. Random forests cannot straightforwardly estimate variance arising from speakers who behave differently, typically incorporated as a random effect in a mixed model (Tomaschek et al., 2018, p. 265), while supervised component regression has low predictive accuracy (Tomaschek et al., p. 260). Moreover, these different approaches do not converge on the same results (Tomaschek et al.).

Figure 1

Correlations between the eight acoustic parameters.

Second, a model with several acoustic parameters as fixed effects should ideally account for speaker-specific variation through random slopes to avoid inflated false positive rates (e.g., Barr et al., 2013). The complexity of such a model would quickly become computationally intractable. The approach of running separate models for each parameter, although not capturing correlation and interactions between the parameters, was therefore chosen. The consequence of this choice is that we cannot isolate the impact of individual acoustic dimensions per se. Instead, the effect magnitude of a parameter represents the magnitude of this parameter along with its correlates. Since we do this for all parameters, they remain comparable.

In the computation of the models, each model ran four MCMC chains for 4000 iterations with 2000 warmup iterations, thus resulting in 8000 posterior samples used for inference. We used regularizing, weakly informative priors for the models (Gelman et al., 2017). For the regression coefficient of interest, we used a normally distributed prior with a mean of 0 and a standard deviation of 0.5. For the intercept, we used a normally distributed prior with a mean of 0 and a standard deviation of 1. For the standard deviation of the random intercepts and slopes and the residual standard deviation, we used a Cauchy distributed prior with a mean of 0 and a standard deviation of 0.5. Rhat values were inspected and indicated convergence of each model, and all models exhibited a high amount of effective samples for all parameters. We therefore concluded that all models were healthy.

In the interpretation of the output of the models, we compared the estimates of the posterior distributions for the two fricatives /ʃ/ and /ç/ for each acoustic parameter. Taking center of gravity as an example, our goal was to investigate whether /ʃ/ had smaller or larger measurements for this parameter than /ç/. We therefore extracted the posterior distributions for /ʃ/ and /ç/ from the model containing center of gravity, and we subtracted the values of the distribution for /ʃ/ from the values of the distribution for /ç/. The mean of the resulting distribution represents the estimated difference in center of gravity between the two fricatives. The 95% Credible Intervals of the distribution were extracted, representing the range of values for which we can be 95% certain we will find the true value of the parameter, given the data, the priors, and the model (Vasishth et al., 2018, p. 152). If the 95% Credible Interval does not contain zero, then, it is likely that there is in fact a difference between /ʃ/ and /ç/ for center of gravity. Importantly, the model estimates also allow us to evaluate effect sizes, which means that we can compare and rank the magnitudes of the differences between /ʃ/ and /ç/ for the different acoustic parameters.

2.3.1. Acoustic results

As the purpose of this study is to explore whether it is possible to characterize speakers in our population as merged and distinct speakers, the most important measurements for the descriptive analysis of the production data are not the recorded values for each acoustic parameter in themselves, but rather the differences between /ʃ/ and /ç/ for each speaker and each acoustic parameter. By studying the distributions of these differences, we can investigate the degree of merging in the population, in that larger differences are indicative of a contrast between the two fricatives, and smaller differences are indicative of a weakening of the contrast.

In order to perform a descriptive analysis of the differences between /ʃ/ and /ç/ for each parameter, speakers’ mean values for the two fricatives were calculated, and the mean value for /ʃ/ was subtracted from the mean value for /ç/. Figure 2 illustrates the mean differences between /ʃ/ and /ç/ for the acoustic parameters measured in Section 2.2.2. A positive difference indicates that /ç/ had a larger mean value than /ʃ/, while a negative difference indicates that /ʃ/ had a larger mean value than /ç/. Density plots were computed using ggplot2 from the Tidyverse package (Wickham et al., 2019) in RStudio (RStudio Team, 2019), and the default values for the smoothing bandwidth and kernel were used.

Figure 2

Distributions of speakers’ mean difference between /ʃ/ and /ç/ for each of the eight measured acoustic parameters. All acoustic parameters were standardized for speaker and vowel. A positive difference indicates that /ç/ had a larger mean value than /ʃ/, while a negative difference indicates that /ʃ/ had a larger mean value than /ç/.

Figure 2 illustrates that for all parameters but duration, a majority of speakers appear to produce a difference between the two fricatives, as indicated by peaks in the distributions left or right of zero. However, we also find a number of speakers close to zero, indicating that certain speakers do not produce a difference between /ʃ/ and /ç/ for these parameters. For duration, on the other hand, a majority of speakers have a mean difference of zero.

Interestingly, speakers’ differences are found on both sides of zero for most of the parameters, indicating that while some speakers have higher values for /ʃ/ than for /ç/ for a given parameter, others have the opposite pattern. This finding could possibly indicate that the different acoustic parameters interact in ways we do not fully understand. For example, speakers could differ in what acoustic cue they emphasize, with the result that their less important cues fluctuate somewhat randomly around zero.

Speakers’ mean differences thus show that for most parameters, we find some speakers who produce a clear difference between /ʃ/ and /ç/, while other speakers produce only a small difference or no difference. There is, however, no clear grouping of speakers. If the speakers in this study formed two distinct groups of merged and distinct speakers, we would expect the distribution of at least some of these parameters to fall into two separate bins, with one group of speakers with a mean difference around zero, and another group of speakers centered on a larger positive or negative value. Since we do not find such a bimodal distribution in the data, we cannot identify merged and distinct speakers based on the measured acoustic differences and, furthermore, we cannot determine which of these acoustic parameters are most important to the production of the contrast between /ʃ/ and /ç/. Consequently, we turn to the statistical analysis investigating which of the measured acoustic parameters contribute more and less to the production of this contrast.

2.3.2. Ranking of cues

The differences between the two fricatives were modeled based on the change from /ʃ/ to /ç/, and the magnitude of the difference for each acoustic parameter was represented by the size of the coefficients for each model. A positive coefficient indicates that the measurements for /ç/ were higher than those for /ʃ/ for the relevant parameter, while a negative coefficient indicates that the measurements for /ʃ/ were higher than those for /ç/. The coefficients were ranked to compare their effect magnitude. We will discuss the effect magnitudes alongside their uncertainty by the posterior means and their 95% Credible Interval. Only a subset of the effect magnitudes are reported here, but posterior means and 95% Credible Intervals for all parameters in all vowel contexts can be found in Appendix A.

Figure 3 illustrates this ranking of coefficients, separated by vowel context. Overall, /ʃ/ and /ç/ seem to be well separated in production, with large effect magnitudes for all vowel contexts. However, as illustrated, the ranking varies substantially depending on the vowel, with /u/ standing out in particular from /i/ and /ɑ/. The results indicate that the separation between /ʃ/ and /ç/ is greatest before the vowel /i/ and smallest before the vowel /u/. While the largest beta weight for /i/ is standard deviation with a size of 1.48 [95% CrI: 1.18, 1.77], the largest beta weight for /u/ is kurtosis with a size of 0.70 [0.39, 1.02]. The effect magnitude for the rest of the parameters are also overall larger before /i/ than before /u/. As for the /ɑ/ context, the largest beta weight is skewness (–0.90 [–1.16, –0.63]), and the effect magnitudes for the remaining parameters appear to fall in between those for /i/ and /u/.

Figure 3

Ranking of the coefficients from the individual regression models, indicating the effect magnitudes of the different parameters. Whiskers indicate 95% Credible Intervals.

Looking at the overall separability of the different acoustic parameters, we find some patterns despite the variation by vowel context. The parameters that appear to give a reliable separation between /ʃ/ and /ç/ across all three vowel contexts are intensity, center of gravity, and kurtosis. The largest effect magnitudes, however, are found for standard deviation and skewness, but only in the /i/ and /ɑ/ contexts. These two parameters do not appear to separate between /ʃ/ and /ç/ in the /u/ context. F2 has a high ranking before /u/ and is found in the middle of the ranking before /ɑ/, although the effect magnitude in the two contexts are similar. Before /i/, however, F2 appears to give no separation. A further discussion of these context effects will follow in Section 4.2. Common to all three vowel contexts is the finding that aspiration and duration are found towards the bottom of the rankings with relatively small beta weights. We can therefore conclude that these parameters do not contribute notably to the separation between /ʃ/ and /ç/ in production.

Having identified certain parameters which contribute more to the contrast than others, we can investigate speakers’ mean differences between /ʃ/ and /ç/ for these parameters specifically. Choosing the four parameters which appear to give the largest degree of separation for the three vowel contexts combined, namely standard deviation, skewness, intensity, and center of gravity, we examined speakers’ mean differences. Figure 4 presents the mean difference for each speaker for these four parameters. As in Figure 2 above, speakers found around zero do not have a notable difference for the relevant acoustic parameter, whereas speakers found to have larger positive or negative values produced a difference for this particular parameter. Figure 4 makes it possible to compare speakers’ differences across parameters, and we can observe that for some speakers, we find some consistencies. Speakers QN and HC, for example, have differences close to zero across these parameters. Speakers, DV and ZV, on the other hand, have large differences for two of these four parameters. However, there is also a lot of variation between parameters, and the distributions form continua, making it impossible to establish non-arbitrary boundaries between speakers.

Figure 4

Individual speakers’ mean difference between /ʃ/ and /ç/ for the four parameters, which seem to contribute most to the separation between these phonemes in production. Examples of speakers with smaller differences (QN, HC) across the different parameters are indicated with light gray dots, and examples of speakers with larger differences (DV, ZV) are indicated with black dots.

We also note that Figure 4 does not take vowel context into account. It is possible that speakers’ mean differences between /ʃ/ and /ç/ for the different parameters are the result of more complex within-speaker patterns depending on the following vowel. These potential patterns are not explored further here, but figures illustrating the rankings of speakers’ mean differences for all acoustic parameters separated by vowel context can be found in Appendix B.

The results of the production study, then, point to certain speakers who have smaller and larger differences, respectively, between /ʃ/ and /ç/. However, despite certain patterns, it is clear that there is no undisputable way to group the speakers in this sample as merged and distinct speakers based on acoustics alone.

3.1.1. Participants

Sixty-four native speakers of Norwegian participated in the experiment (49 female, 15 male; aged 18–50, mean age 24). As in the production study, participants were first-year students of Scandinavian studies completing an introductory course in Norwegian grammar, and they received course credits for participation. The criteria for inclusion were less strict in this study than in the production study. As well as East Norwegian speakers, speakers with other Norwegian dialects and other native languages in addition to Norwegian were included. As East Norwegian constitutes the standard spoken language in Norway, all participants have been extensively exposed to this variety throughout their lifetime, and they all lived in Oslo at the time of participation. It is therefore reasonable to assume that their perception of /ʃ/ and /ç/ is comparable to the perception of East Norwegian speakers.³

Furthermore, participants were asked to report whether they had any speech or hearing impairments. One participant reported to have a speech impairment, but because only perception was under investigation in the current experiment, this participant was nonetheless included. None of the participants reported to have a hearing impairment.

3.1.2. Materials

The experiment consisted of a practice phase and a test phase, and the materials used in each phase were different. In the practice phase, the materials consisted of nonce words produced by three native speakers of East Norwegian who did not participate in the production study (2 female, 1 male; aged 30–42). The nonce words were identical to the nonce words containing /ʃ/ and /ç/ in the production study, giving 6 words with a CV structure, where C was /ʃ/ or /ç/ and V was /i/, /ɑ/, or /u/. Each speaker produced these 6 nonce words once, giving 18 words in total in the practice phase. The speakers were aware of the purpose of the experiment, and they were recorded under the same conditions as the participants in the production study. The purpose of including this practice phase was to allow the listeners to familiarize themselves with the task they were going to complete.

In the test phase, the recorded nonce words containing /ʃ/ and /ç/ from the production study were used as stimuli. To ensure that an equal number of productions of each fricative was included for each speaker, the speakers who occasionally produced [sj] and [kj] for /ʃ/ and /ç/ in the production study (see Section 2.2.1) were excluded. This left 34 speakers whose productions were included as stimuli in the perception study. The materials consisted of 6 words, repeated twice, for each of the 34 speakers, resulting in 408 tokens in total.

3.1.3. Procedure

The experiment took place in the Sociocognitive Lab at the University of Oslo, and the data were collected through a Praat MFC Experiment (Boersma & Weenink, 2022). Participants were seated in front of a laptop computer and informed that they would be participating in a listening experiment. As mentioned above, the experiment included a practice phase of 18 trials and a test phase of 408 trials. In both phases, the nonce words were presented in a randomized order. The test phase contained 6 blocks of 68 trials, with breaks between each block.

Participants were instructed that they would hear a series of words and that their task was to choose which sound each word started with. They were told that the first sound would always be one of two sounds: the sj-sound, corresponding to /ʃ/ or the kj-sound, corresponding to /ç/. As the participants were not necessarily familiar with the phonetic symbols for these sounds, the orthographic representations ‹sj› and ‹kj› were used when displaying the two options on the screen (see Section 2.1.3).

To choose between the two sounds, participants were told to press the S key on the laptop keyboard for the kj-sound and the L key for the sj-sound. Participants could also press the space-bar to replay a word. This option was included in case participants did not hear a given word, either because of noise from the environment or because they lost focus. Participants were informed that this option was intended for cases such as the ones mentioned, rather than to repeat every word they were unsure of.⁴

Participants wore Koss UR40 over ear headphones during the experiment, and they were told to adjust the volume to a comfortable level. The experiment lasted approximately 25 minutes.

3.3.1. Identification

Starting with the level of correct identification, Figure 5 illustrates the distributions of listeners’ identification scores. That is, the proportion of correctly identified tokens for each listener, for each of the three vowel contexts. The error bars represent the posterior means and the 95% Credible Intervals for correct identification, as estimated in the logistic regression model described in Section 3.2. The posterior means and 95% Credible Intervals in the /i/ context (0.75 [0.71, 0.80]), the /ɑ/ context (0.76 [0.72, 0.80]), and the /u/ context (0.77 [0.72, 0.81]) indicate that, based on the model, the data, and the priors, it is plausible that listeners as a population were above chance in identifying speakers’ productions in all vowel contexts. Furthermore, as the 95% Credible Intervals largely overlap for the three vowel contexts, it is plausible that there was no difference in identification of /ʃ/ and /ç/ depending on the following vowel.

Figure 5

Proportions of correctly identified responses for listeners as a population in the three vowel contexts /i/, /ɑ/, and /u/. Individual dots are listener averages. Density represents smoothed kernel density of these averages. Shading corresponds to the estimated middle 50 / 80 / 95% of the distributions. Point corresponds to posterior means. Error bars correspond to the 95% Credible Intervals around the posterior means.

Overall, then, /ʃ/ and /ç/ are well separated in perception. Note, however, that the distributions overlap with an area around chance performance. In other words, while the majority of listeners perceived the fricatives above chance, the lower tails of the distributions contain listeners who appear to have guessed which fricative to respond, indicating that they did not accurately perceive the fricatives.

3.3.2. Accuracy

We now turn to an investigation of the level of correct identification by speaker, a measure we will refer to as accuracy. This measure demonstrates to what extent listeners perceived intended /ʃ/ as /ʃ/ and intended /ç/ as /ç/ for a given speaker, giving an indication as to the degree of merging among speakers as seen from a perceptual perspective. Imagine, for example, that listeners correctly identified 75% of a speaker’s productions of /ʃ/ and 70% of the same speaker’s productions of /ç/. This speaker’s overall accuracy score would be the mean of these two scores, namely 72.5%. Speakers with high accuracy scores of 90%, for example, are likely to have produced a contrast, as the level of accuracy is high for both fricatives. Speakers with accuracy scores around 50%, on the other hand, most likely have not produced a perceptible contrast. This score could be the result of different underlying patterns.

One type of speaker could have a high accuracy score for one of the fricatives and a low accuracy score for the other. For instance, if, for a given speaker, listeners identified intended /ʃ/ as /ʃ/ 80% of the time, but in contrast identified intended /ç/ as /ç/ 20% of the time, the total accuracy score for this speaker would be 50%. This score reflects a scenario where the speaker has a merged pronunciation, and the result of the merger is /ʃ/. That is, intended /ʃ/ is perceived as /ʃ/, and intended /ç/ is also perceived as /ʃ/.

Another imaginable scenario is that a speaker has an overall accuracy score of 50%, just as in the previous scenario, but that the accuracy score for both fricatives is around 50%. In this case, listeners were at chance when identifying whether the speaker produced /ʃ/ or /ç/, indicating that the speaker’s productions of both fricatives were not easily categorized as either /ʃ/ or /ç/. This speaker is likely to have a merged pronunciation, but the result of the merger might be somewhere in between the two fricatives.

Figure 6 illustrates the distributions of accuracy scores for the population of speakers, separated by vowel context. As in Figure 5, the error bars indicate the posterior means and 95% Credible Intervals for correct identification based on the Bayesian logistic regression model described in Section 3.2 (0.75 [0.71, 0.80] in the /i/ context, 0.76 [0.72, 0.80] in the /ɑ/ context, and 0.77 [0.72, 0.81] in the /u/ context). It becomes clear that a majority of speakers have accuracy scores well above 50% in all vowel contexts, indicating that they produced a perceivable contrast between /ʃ/ and /ç/. A smaller number of speakers, however, are found in an area around 50%, indicating that listeners did not reliably perceive a contrast between these speakers’ fricative productions.

Figure 6

Proportion of correctly identified responses for speakers as a population, separated by vowel context. Individual dots are speaker averages. Density represents smoothed kernel density of these averages. Shading corresponds to the estimated middle 50 / 80 / 95% of that distribution. Point corresponds to posterior mean. Error bar corresponds to the 95% Credible Interval around the posterior mean.

Furthermore, Figure 7 presents the accuracy scores of individual speakers, collapsed across vowel contexts as the Bayesian logistic regression model, and Figures 5 and 6 all indicated that there were no notable differences in the degree of correct identification of /ʃ/ and /ç/ depending on the following vowel. Figure 7 illustrates that speakers form a continuum, from scores around 55% to scores above 80%. It is clear from the distribution that there is no natural cutoff point between speakers which would allow us to single out separate groups. Moreover, Figure 7 also presents speakers’ accuracy scores for the two fricatives separately, and it is clear that among the speakers with the highest overall accuracy scores, there does not appear to be a notable difference between /ʃ/ and /ç/. Among the speakers with the lowest accuracy scores, we find both types of speakers described above. Looking at the five speakers with the lowest scores, GX, FT, and QN have a relatively high score for one fricative and a low score for the other, while HC and UQ have relatively low scores for both fricatives. Note that GX’s scores indicate a merged pronunciation in the direction of /ʃ/, while FT and QN’s scores are indicative of a merged pronunciation in the direction of /ç/.

Figure 7

Proportion of correctly identified responses for each speaker. Proportions for the individual fricatives are indicated by their IPA symbols.

Figure 7 also shows that no speaker has a higher overall accuracy score than 83%. If we take the findings in the identification analysis in Section 3.3.1 into account, it is reasonable to assume that speakers’ scores are affected by the fact that some listeners were unable to correctly identify the fricatives (see Figure 5). These listeners were most likely at chance performance for all speakers, and their effect on the accuracy scores is therefore expected to be the same across speakers. It is probable, then, that accuracy scores would be higher overall if it were not for these listeners.

We therefore carried out a descriptive analysis of the accuracy scores for individual speakers based only on the perception of the listeners who most likely were able to perceive the contrast between /ʃ/ and /ç/. We used 75% correct identification of fricative productions as a pragmatic threshold for having perceived the contrast (see Treutwein, 1995; Leek, 2001; Lesmes et al., 2015, for references to the use of 75% as a target performance level for the identification of stimuli in psychological and forced-choice experiments). This selection resulted in 34 listeners, who will be referred to as the top listeners.

As illustrated in Figure 8, accuracy scores are markedly higher when only top listeners are included. A majority of speakers have scores above 75%, and the highest ranked speakers have scores up to 93%. As predicted, then, the accuracy scores in Figure 7 were lowered by the inclusion of listeners who did not reliably perceive the contrast between /ʃ/ and /ç/. Note, however, that the increase in accuracy scores in Figure 8 compared to Figure 7 is clearly smaller for the four lowest ranked speakers than for the rest of the speakers. The two speakers at the bottom, GX and FT, hardly improve at all, while the next two speakers in the ranking, QN and HC, show only slight improvements. The next two speakers, UQ and SC, on the other hand, show considerable improvements, along with the rest of the speakers.

Figure 8

Proportion of correctly identified responses for each speaker based on the perception of top listeners only. Proportions for the individual fricatives are indicated by their IPA symbols.

The comparison between Figures 7 and 8 thus indicates that the four speakers at the bottom of the ranking likely did not produce a contrast between /ʃ/ and /ç/. That is, if these speakers in fact produce no difference between the two fricatives, their accuracy scores should not be much higher for top listeners than for all listeners. Even if listeners can reliably perceive the contrast, they cannot perceive a distinction which the speaker did not produce.

As in Figure 7, Figure 8 presents speakers’ accuracy scores for the individual fricatives as well as their mean accuracy score. These scores largely corroborate the scores in Figure 7.

3.3.3. Ranking of cues

The results of the identification task presented in sections 3.3.1 and 3.3.2 illustrate the extent to which listeners accurately categorized speakers’ fricative productions. We now turn to an investigation of which acoustic parameters listeners used when they performed this categorization. By studying the contribution of the different acoustic parameters measured in the production study to the perception of the contrast between /ʃ/ and /ç/, similar to the ranking of parameters performed in Section 2.2.3, we can compare whether speakers and listeners rely on the same parameters.

Figure 9 presents the log odds of the logistic regression models for each acoustic parameter, separated by vowel context and ranked by size. As was the case in the corresponding analysis in Section 2.2.3, the differences between /ʃ/ and /ç/ were modeled as the change from /ʃ/ to /ç/. Again, positive coefficients indicate that the measurements for /ç/ were higher than those for /ʃ/, and negative coefficients indicate that the measurements for /ʃ/ were higher than those for /ç/. Effect magnitudes are reported alongside their uncertainty by the posterior means and their 95% Credible Interval. Only a subset of the effect magnitudes are reported here, but posterior means and 95% Credible Intervals for all parameters in all vowel contexts can be found in Appendix C.

Figure 9

Ranking of the coefficients from the individual regression models, indicating the effect magnitudes of the different parameters. Whiskers indicate 95% Credible Intervals.

Figure 9 illustrates that the fricatives seem to be well separated in perception, as posterior means and 95% Credible Intervals do not include zero as a plausible value for most of the parameters. As in the production study, however, the ranking of parameters shows variation by vowel context. The highest ranked parameter in the /i/ context is kurtosis (–1.01 [–1.20, –0.81]), while in the /ɑ/ context, it is intensity (–1.11 [–1.35, –0.88]), and in the /u/ context F2 (1.09 [0.96, 1.23]). Overall, the /ɑ/ context seems to give the best separation between /ʃ/ and /ç/, while /i/ and /u/ have similar effect magnitudes. These vowel context effects will be discussed further in Section 4.2.

Assessing the overall ranking of the different parameters, intensity appears to be the parameter which gives the largest degree of separation between /ʃ/ and /ç/ in perception, as it is ranked highly for all vowel contexts. Standard deviation, center of gravity, and skewness also have a fairly high ranking across vowel contexts, although skewness has a low ranking before /u/. Duration and aspiration are consistently ranked towards the bottom of the list, indicating that these parameters do not contribute notably to the separation between /ʃ/ and /ç/ in perception. The remaining parameters are more variable depending on the vowel context. F2 is ranked highly before /ɑ/ and /u/, but it appears to be of no relevance before /i/. Conversely, kurtosis is the highest ranked parameter before /i/, while it is the lowest ranked parameter before /u/. Before /ɑ/, it is found in the middle of the ranking.