36 participants from three occupational groups took part in the investigation. The occupations were chosen due to their traditional association as either male dominated (soldiers working in military barracks, persons in leading positions) or female dominated (kindergarten teachers). The two male dominated occupation groups were not combined, due to their differing socio-economic characteristics as well as differences in prestige. Participants had to have at least 3 years of experience in their respective occupations to ensure a base level of identification with their job.

Table 1 summarizes participants’ age, sexual orientation, years of practice and identification with occupation separated by gender and occupation. Participants were asked to identify their gender as female, male or diverse. As participants chose either female or male, in what follows we will use female and male (in contrast to AFAB and AMAB) with respect to our subjects. Participants rated their identification with their job—on a scale from not at all (1) to very much (5)—from medium to high (at least 3.3). They had several years of experience in their respective jobs. Sexual orientation was assessed using a Kinsey-like scale (Kinsey et al., 1948) ranging from exclusively heterosexual (1) to exclusively gay/lesbian (7) or asexual (9) to account for potential differences in fine phonetic detail based on sexual orientation (Munson et al., 2006). While most of the participants chose scores leaning towards the heterosexual pole, there was some variation regarding the exclusivity of being heterosexual. Two participants chose 4 and two chose either 5 or 7. Importantly, there is no systematic difference in sexual orientation based on our factor of interest, i.e., the occupation group.

To investigate participants’ gender expression and to look at potential interactions of gender-related personality dimensions with occupation groups and phonetic variation, three questionnaires were used (chosen based on prior socio-phonetic studies (Kachel et al., 2017; Weirich & Simpson, 2018; Weirich & Zahner-Ritter, 2023). The Normative Gender Role Orientation scale (NGRO, Athenstaedt, 2000) ranging from I fully agree (1) to I do not agree (7) was used to assess normative gender role orientation. Femininity (expressivity) and masculinity (instrumentality) was measured using the respective scales F+ (positive attributes traditionally attributed to females) and M+ (positive attributes traditionally attributed to males) of the GEPAQ (Runge et al., 1981) on a scale from, e.g., dependent (1) to independent (7). Thirdly, identification on the Traditional-Masculinity-Femininity-Scale (TMF) regarding traditional male/female roles (Kachel et al., 2016) was assessed on a scale from very masculine (1) to very feminine (7).

Cronbach’s α of the four scales was medium to high, with the lowest value for NGRO (α =.70), followed by GEPAQ_M (α = .80), GEPAQ_F (α = .85) and TMF (α = .96). Figure S1 visualizes the correlation matrix of the gender-scales (significant correlations with p < .05 marked with an asterisks). The femininity and masculinity scales of GEPAQ did not correlate, mirroring the independence of these two dimensions. TMF was positively correlated with GEPAQ_F+ (with traditionally more female ratings on TMF corresponding to higher femininity), while GEPAQ_M+ seems to capture other aspects than TMF. NGRO correlated positively with GEPAQ_F+ and negatively with GEPAQ_M+.

For the f0 analysis, 3 female kindergarten teachers were excluded due to their heavy smoking behavior (> 10 cigarettes per day). In combination with their age (> 57) this resulted in very low f0 values (130Hz, 150Hz, 170Hz). Age and smoking behavior do not affect formants in the same way, so all 36 participants were included in the formant analysis, while 33 participants (16 females and 17 males) were included in the f0 analysis (see Table S1, Weirich, 2025a).

Participants were recorded in three cities in East Central Germany (Jena, Erfurt, Dresden). The speech material consisted of two tasks, 1) a reading task with 15 sentences which were repeated twice and 2) a dialogue task, where participants were paired within occupations and given two (out of 6) pictures that varied slightly (see Figure S2, Weirich, 2025a). Since it has been shown that speakers adapt to interlocutors and situations (Bell, 1984; Giles et al., 1991), participants were recorded in their respective occupational role speaking to a colleague (in the kindergarten, in the offices of the leaders, in the military barracks).

The sentences of the reading task contained target words particularly suitable to measure formant values in peripheral vowel qualities in stressed syllables (such as /i:/ in Biene (bee), /a/ in Tasse (cup), /u:/ in Luka (first name)). For the dialogue task, custom-made pairs of pictures were used showing objects and animals with vowels in stressed syllables (e.g., /i:/ in Liege (lounger), /a/ in Tasche (bag), /u:/ in Kuh (cow), see Table S2 for all included vowels and words, Weirich, 2025a). Participants were asked to find differences between scenes by talking about their pictures (“Spot the difference”). Thus, the conversation between participants was a task-oriented dialogue in a game form to encourage free communication, often used in studies on accommodation (Pardo, 2006). All recordings were made using a headset microphone (DPA, CORE Omni Headset) connected to a ZOOM–H6 Handy Recorder, making it possible to have separate high quality audio signals for the participants necessary for the acoustic analyses. The dialogues lasted on average 11.75 min (SD = 5.6 min).

Prior to acoustic analyses, the speech data was orthographically transcribed and further segmented automatically into words and segments using MAUS via its online interface WebMAUS Basic (Kisler et al., 2017). All acoustic analyses were performed using PRAAT (Boersma & Weenink, 2015). Mean f0 and standard deviation (SD) of mean f0 were measured for each participant separately for the two speech tasks²2

analysis parameters: time step = 0.01; minimum pitch for males = 50 Hz, for females = 75 Hz; maximum pitch for males = 350 Hz, for females = 450 Hz

For read speech, only one value per parameter was used for each speaker (the average f0 estimated across the reading task (=mean f0) and the standard variation of the f0 measurements across this task (=var f0)).

Formant frequencies (F1, F2) were measured at the midpoint of the labeled vowels using PRAAT’s LPC formant measurement algorithm³3

analysis parameters: time step = 0.01s; maximum number of formants = 5; window length = 0.025s; preemphasis from = 50 Hz; maximum formant value for males = 5000 Hz, for females = 5500 Hz

For the dialogue task, formants (F1 and F2) in 2,145 vowel tokens (1,102 from female speakers) were measured. Analogous to the f0 analysis, each participant was included only once. Due to the nature of the speech material (a spontaneous dialogue is less controlled than read speech and contains more speech errors, hesitations, overlapping speech) more formants were incorrectly measured than in the read speech and, thus, were not manually corrected but removed from the data set as outliers. For each vowel and formant, the interquartile range (IQR) (difference between the 75th percentile (Q3) and the 25th percentile (Q1)) was measured. An observation was considered to be an outlier if it was 1.5 times the interquartile range greater than the third quartile (Q3) or 1.5 times the interquartile range less than the first quartile (Q1). After outlier exclusion, 1,760 vowel tokens were subjected to the analysis.

For statistical analyses, linear (mixed) models in the R environment (R Core Team, 2020) were run. Test variables were added successively to the model using a forward stepwise regression. p-values were estimated using model comparisons, i.e., by comparing models with and without the respective variables or interactions in question using likelihood ratio tests provided by the anova function (package lme4, Bates et al., 2015). Summaries of the final models are given in the Supplementary Materials (see Tables S4–S23, Weirich, 2025a). For post hoc comparisons, the package emmeans was used (Lenth, 2022). Pearson correlations were run after testing for normal distributions (Shapiro-Wilk normality tests, all p-values > .10) of the respective variables (e.g., mean f0, GEPAQ_F+).

Table 2 shows mean values and SDs of the gender variables separated by gender and occupation. The numbers only point to clear gender differences for TMF, while all other scales show similar values between the groups or do not differ with respect to gender but rather to occupation (higher masculinity in male leaders and in male soldiers than in female leaders and female soldiers but lower masculinity in male kindergarten teachers than female kindergarten teachers). Accordingly, the linear model revealed a significant effect of GENDER for TMF, F(1) = 87.12, p < .001, and also—with a much smaller effect size—for GEPAQ_F+, F(1) = 7.4, p = .01, but not for GEPAQ_M+, F(1) = 0.9, p = .35, nor for NGRO, albeit marginal, F(1) = 3.4, p = .07. While the table reveals job specific variation in gender differences in GEPAQ_F+ (higher femininity in female soldiers and teachers compared to their male counterparts but lower femininity in female leaders than male leaders), including the interaction of GENDER*OCCUPATION into the model failed to reach significance, F(2) = 2.1, p = .099. For GEPAQ_M+ the table reveals lower values for teachers compared to the other groups, but the factor OCCUPATION marginally failed to reach significance F(2) = 3.1, p = .056. For NGRO and TMF, including neither OCCUPATION as a main factor nor the interaction of OCCUPATION *GENDER revealed a better fit (p > .35).

Summary statistics with f0 values averaged across occupation group and task are given in Table S3 (see Weirich, 2025a). Separate linear (mixed) models were run for the speech tasks with either MEAN f0 or VAR f0 as dependent variable. For MEAN f0 and the dialogue task, the LMM (with SPEAKER and SEQUENCE as random intercepts) revealed the expected effect of GENDER, χ²(1) = 61.7, p < .001, while adding OCCUPATION did not result in a better fit of the model (p > 0.7). Similarly, for VAR f0 and the dialogue task, the LMM (with SPEAKER and SEQUENCE as random intercepts) shows an effect of GENDER, χ²(1) = 28.01, p < .001. The interaction of GENDER *OCCUPATION failed to show significance, χ²(4) = 7.23, p = .12.

For read speech, linear models were run (since only one f0 measurement per speaker was measured). For MEAN f0, a significant effect of GENDER was found, F(1) = 138, p < .001, while adding OCCUPATION did not lead to a better fit (p > .8). For VAR f0, both GENDER, F(1) = 9.49, p = .004, and OCCUPATION, F(2) = 4.54, p = .019, showed a significant effect. Post hoc comparisons revealed significant differences between leaders and kindergarten teachers, mean L = 35.7 (SE = 2.57), mean K = 26.4 (SE = 1.192), Estimate = -9.29, df = 27, p = .019, with leaders showing higher variation than teachers. All other comparisons were not significant (p > .17).

Table 3 gives a summary of the f0 measurements separated by speaker and dialogue condition (DG, SG). The linear mixed model with SPEAKER and SEQUENCE as random factors and MEAN f0 as the dependent variable revealed a significant interaction of SPEAKER GENDER and the GENDER of the DIALOGUE PARTNER, χ²(2) = 6.18, p = .045, with females having higher mean f0 values in the DG condition than in the SG condition and males having lower mean f0 values in the DG condition than in the SG condition. This is reflected in the values presented in Table 4 (except for one speaker: M03). Thus, the contrast between the genders in mean f0 is increased in the DG condition. For VAR f0, the interaction marginally failed to reach significance, χ²(2) = 4.61, p = .099, but the numbers in Table 3 also show a gender specific pattern: while females show a decrease in variation in f0 in the DG pairs, males show a slight increase in variation in f0.

To investigate the relationship between self-rated femininity (GEPAQ_F) and f0 parameters, mean f0 and mean variation in f0 were estimated for each speaker from the reading task. Pearson correlations were run between the f0 measures and GEPAQ_F+ scores for males and females separately. For males, a significant positive correlation was found between FEMININTY and MEAN f0, with speakers who rate themselves as more feminine also having a higher mean f0 (r = .49, p = .024, df = 15). In addition, a significant correlation was found between VAR f0 and FEMININTY, with more feminine men having a higher variation in f0 (r = 0.53, p = .014, df = 15). For females, no such relationship was found, neither for mean f0 (r = .15, p = .29, df = 14) nor for variation in f0 (r = -.19, p = .768, df = 14). Figure 1 visualizes the relationships, also giving information about the occupation of the respective speaker.

We also investigated the relationship between f0 parameters and masculinity (GEPAQ_M), but no significant correlation was found (all p > .22). Since both NGRO and TMF correlated positively with GEPAQ_F we did not run further correlations.

Linear mixed models with SPEAKER and WORD as random intercepts were run separately for F1 and F2. For read speech and F2, a significant interaction of VOWEL*GENDER*OCCUPATION was found (χ²(11) = 198.96, p < .001). Post hoc comparisons were run to detect for which gender and vowels occupation groups differ (since we are not interested in the expected differences between vowels and genders, but instead in any occupation specific productions). Significantly higher F2 values were found for female kindergarten teachers than for female soldiers for /e:/, mean K = 2376 (SE = 72.3), mean S = 2156 (SE = 87.2), Estimate =220.5, df = 60.4, p = .002, and also for /i:/, mean K = 2486 (SE = 51.3), mean S = 2305 (SE = 65.0), Estimate = -180.9, df = 39.7, p = .007. In addition, for /e:/, female kindergarten teachers also showed higher F2 values than female leaders, mean K = 2376 (SE = 72.3), mean L = 2207 (SE = 83.4), Estimate = 168.9, df = 56.7, p = .011. For males, kindergarten teachers differed from soldiers by showing higher F2 values for /a/, mean K = 1368 (SE = 48.3), mean S = 1252 (SE = 46.5), Estimate = 115.7, df = 35.1, p = .047. To sum up, the group of kindergarten teachers stood out in terms of higher F2 values than the other groups (mediated by vowel and gender).

For read speech and F1, again a significant interaction between VOWEL*GENDER*OCCUPATION was found, χ²(11) = 48.2, p < .001. Post hoc comparisons revealed significant differences between female kindergarten teachers and female leaders for /a/, with leaders showing higher F1 values than kindergarten teachers, mean L = 811 (SE = 19.7), mean K = 751 (SE =14.4), Estimate = 60.18, SE = 19.5, df = 37.3, p = .01. All other comparisons between occupations had p-values larger than .16.

For the dialogue task and F2—similar to the read speech data—the LMM with the interaction of OCCUPATION*GENDER*VOWEL explained the data best, χ²(11) = 108.49, p < .001. Similar to read speech, the female kindergarten teachers stood out by having higher F2 values for /e:/ than soldiers, mean K = 2449 (SE = 43), mean S = 2231 (SE = 61.5), Estimate = -217.42, p < .001, and also leaders, mean L = 2290, Estimate = -158.57, p = .01. In males, a tendency was found for kindergarten teachers to show higher F2 values than soldiers (p = 0.16).

Also for F1, a significant interaction of VOWEL*GENDER*OCCUPATION was found, χ²(19) = 121.12, p < .001. In male speakers, the leaders exhibited higher F1 values for /a/ than soldiers, mean L = 652 (SE = 16.5), mean S = 600 (SE = 13.7), Estimate = 52.5, p = .022, and also tendentially higher than kindergarten teachers, mean K = 604 (SE=17.2), Estimate = 48.5, p = .078. There was also a tendency for female soldiers to show higher F1 values than kindergarten teachers for /e:/, mean S = 399 (SE = 18.9), mean K = 354 (SE = 12.9), Estimate = 45.3, p = .055.

To sum up, for both read and dialogue speech, male leaders had higher F1 values than the other occupation groups, while for F2 female kindergarten teachers exhibited higher values.

Analogous to the f0 analysis, we looked for an effect of interlocutor’s gender on F1 and F2. LMMs with SPEAKER and WORD as random intercepts were run. For both, F1 and F2 the expected significant interaction of VOWEL*GENDER was found, F1: χ²(3) = 108.8, p < .001, F2: χ²(3) = 54.23, p < .001, indicating phone-specific differences between male and female speakers. In contrast to f0, adding the factor of the INTERLOCUTORs GENDER (same vs. different gender) did not reveal a better fit of the data (p-values > .37).

To explore the relationship between self-rated femininity (GEPAQ_F) and formants (F1, F2) measurements from the reading task were used for each speaker. LMMs were run for male and female speakers separately with either F1 or F2 as dependent variable and WORD and SPEAKER as random effects.

For male speakers and both F1 and F2, the model with the interaction of VOWEL*GEPAQ_F explained the data better than the model with VOWEL alone, F1: χ²(4) = 12.476, p = .014; F2: χ²(4) = 12.95, p = .011. Figure 2 visualizes the effect of femininity on F1 (left plot) and F2 (middle plot) separated by vowel: higher femininity is reflected in higher F1 values for /a/, which corresponds to a lower tongue and jaw position and thus a clearer pronunciation of /a/, while for the high vowels /e:/, /i:/ and /u:/ the direction is the other way, also pointing to a larger vowel space and thus a clearer speaking style with higher femininty values. In addition, higher femininity is reflected in higher F2 values, especially for the front vowels /e:/ and /i:/ but also for /a/, while for the back vowel /u:/ the direction is the other way. Higher F2 values in front vowels and lower F2 values in back vowels correspond to a larger vowel space and thus a clearer speaking style. For both formants, higher femininity ratings are reflected in values shifted towards more female speech (higher formants in females than in males, clearer speaking style in females).

For female speakers and F1, the expected significant effect of VOWEL was found, χ²(3) = 73.85, p < .001, but adding GEPAQ_F to the model did not result in a better fit (p > 0.5). For female speakers and F2, the interaction of VOWEL*GEPAQ_F turned out to be significant, χ²(4) = 18.15, p = 0.001: while F2 for /i:/ decreases with higher femininity, F2 increases with higher femininity for /u:/. This contrasts with the male patterns and the expected relationship between femininity and clear speech.

130 participants took part in the listening experiment which was run using the online platform SosciSurvey (Leiner, 2018) and lasted approximately 15 minutes. Participants over a wide age range were recruited in university classes, over social media as well as through diverse contacts of the author and colleagues (e.g., family, friends, sports clubs, orchestra, etc.).

Participants were given information about the supposed aim of the study, i.e., investigating personality attributions based on voices. Crucially, descriptions differed in the priming of the listeners regarding the supposed speakers: Listener group 1 was told that recordings were made in a kindergarten, while listener group 2 was told that recordings were made in military barracks. In addition, for each stimulus to be rated in the experiment the question was prefaced with “Please rate the kindergarten teacher” for the first listener group and “Please rate the soldier” for the second listener group. Note that in this priming experiment, the leader category was not used. Listeners were told that there were no right, or wrong answers and they should decide intuitively. In addition, participants were asked about their age, mother tongue, where they lived, educational background, gender (by choosing female, male, non-binary or no answer) and sexual orientation (to be rated on a Kinsey-like scale).

Table 4 gives a summary of participants’ background separated by priming condition and age group. In a previous study (Weirich et al., 2020), listeners’ age was found to be a relevant factor regarding stereotypical assumptions about speaker groups. Therefore, age was added as a potential factor exploratorily investigating differences in the prestige of occupational groups.

A pretest was run to familiarize participants with the experimental design. After the main experiment participants were thanked and the true aim of the investigation was explained.

Acoustic stimuli came from 10 males (5 soldiers, 5 kindergarten teachers) and 10 females (3 soldiers, 7 kindergarten teacher) and consisted of the sentence “Der Hase frisst die Möhre” (The rabbit eats the carrot). These stimuli were part of the reading task elicited in the production study. By using sentences as stimuli, listeners were provided with both segmental cues (such as vowel formants, sibilant spectra) and supra-segmental information (such as mean f0, variation in f0, speaking tempo) known to be used for speaker attributions such as gender, femininity, or personality (Klofstad et al., 2012; Weirich & Simpson, 2018).

The procedure of the perception test was the same for both priming groups and each stimulus. A text was presented asking the participant to rate the (assumed) kindergarten teacher (priming group = K) or the (assumed) soldier (priming group = S) based on their voice. Participants could listen to each stimulus several times but were still asked to rate fast and intuitively. 8 bipolar item pairs were provided with the possibility to rate the stimulus regarding each of these item pairs on a 7-point scale. The items were chosen to reflect the personality dimensions warmth⁴4

Consisting of the items warm – cold, likeable – unlikeable, sensitive – insensitive, emotional – unemotional

Cronbach’s alpha was measured to estimate reliability for the items used for the personality dimensions warmth and competence resulting in high values for both scales (.88 and .83 respectively). Mean values for the two dimensions were calculated across all respective items and used in linear mixed models as dependent variables using the R environment (R Core Team, 2020). Test variables were added successively to the model using a forward stepwise regression. Summaries of the final models are given in the Supplementary Materials (see Tables S24–S25, Weirich, 2025a). p-values were estimated using model comparisons, i.e., by comparing models with and without the respective variables or interactions in question using likelihood ratio tests (lme4, Bates et al., 2015). LISTENER and SPEAKER were included as random intercepts. We tested for the effects of PRIMING, AGE GROUP, LISTENER GENDER and SPEAKER GENDER. For post hoc comparisons, the package emmeans was used (Lenth, 2022).

For warmth, a significant effect of LISTENER GENDER, χ²(3) = 11.5, p = 0.009, and an interaction of PRIMING *AGE GROUP, χ²(3) = 17.5, p < .001, was found. Female listeners rated the speakers to be warmer than male listeners (Estimate = -0.22652, SE = 0. 0675, df = 171.23, p < .001), independent of age group, speaker gender and priming. Furthermore, younger listeners rated supposed kindergarten teachers to be warmer than supposed soldiers, mean K = 4.51 (SE = 0.20), mean S = 4.19 (SE = 0.19), Estimate = 0.321, p = .002, while older listeners rated supposed soldiers to be warmer than supposed kindergarten teachers, mean K = 4.16 (SE = 0.19), mean S = 4.34 (SE = 0.19), Estimate = -0.182, p = .025, see Figure 3 (left plot).

For competence, no effect of priming or age group was found, but the speakers’ gender and the listeners’ gender affected the ratings. Here, we excluded the two non-binary listeners due to their low number in comparison to the female and male participants. The model with the best fit of the data included the interaction of SPEAKER*GENDER*LISTENER GENDER, χ²(3) = 11.01, p = .012, see Figure 3 (right plot). Male speakers were more positively rated in terms of higher competence by female listeners in comparison to male listeners, mean female listeners = 4.49 (SE = 0.17), mean male listeners = 4.25 (SE = 0.17), Estimate: 0.236, p = .002. For female speakers, the gender of the listener had no significant effect, mean female listeners = 4.55 (SE = 20.9), mean male listeners = 4.48 (SE = 21.6), Estimate = 0.078, p = .289.