Female Budgerigars Attend Equally to Rhythmic Beat or Repetition, While Humans Prioritize Their Combination

2.1.1. Subjects

We tested 13 adult budgerigars (seven females, six males), aged between 1 and 4 years. Birds were housed in the Animal Care Facility at the Department of Behavioral and Cognitive Biology, University of Vienna, in three mixed‐sex aviaries (two of size 2 × 1 × 2 m, one of size 2 × 1.5 × 2 m), each containing groups of 5–10 birds. A 12.5/11.5 h light/dark cycle was maintained, approximating their natural diurnal cycle [23]. Water, periodically supplemented with vitamin supplements and bird food pellets, was available ad libitum. None of the birds had prior experience with the testing apparatus, and the birds were naive to the current stimulus set. However, 10 of the birds (six females, four males) had participated in the Hoeschele and Bowling [20] study and, therefore, had experience with a different two‐alternative forced‐choice preference apparatus. All birds were accustomed to food‐rewarded operant conditioning tasks testing auditory cognition (during the period of the current studies: lexical stress [24] and octave equivalence [25]). All procedures involving animals in the studies described here were conducted in accordance with Austrian animal protection and housing laws and were approved by the ethical board of the behavioral research group in the faculty of Life Sciences at the University of Vienna (Approval Number 2015–005).

2.1.2. Apparatus

Budgerigars were tested in a custom‐built hexagonal aviary that was acoustically isolated from the main aviaries in which they were normally housed. This testing aviary comprised three perches, each paired with an adjacent FE108 full‐range speaker (Fostex, Tokyo, Japan; frequency range: 200–16,000 Hz). Landing on a perch disrupted an infrared beam (IR Break Beam Sensor—3 mm LEDs, Adafruit Industries, New York, NY, USA), triggering stimulus playback from the speaker after 300 ms of beam disruption (according to a custom written Python script running on a MacBook Pro [Apple Inc., Cupertino, CA, USA] with an Arduino Uno REV 3 chip [BCMI, Turin, Italy] interface). Because pilot testing showed that birds would not interact with the apparatus when acoustically isolated from their flock, we played back 2 h recordings from their respective flock, at an intensity of approximately 50 dB SPL (measured at the bird's position) during testing sessions, using an AV 40 speaker placed above the testing apparatus (M‐Audio, Cumberland, RI, USA) controlled by an iPad (Apple Inc.). Water and food were available ad libitum on the apparatus’ floor throughout the experiment, not administered as rewards.

2.1.3. Stimuli

We created two sets of stimuli—fully rhythmic and arrhythmic—each set consisting of 600 stimuli, with a duration of 60 s per stimulus. Each stimulus was composed of five different budgerigar vocal units (∼100–200 ms in length), extracted from audio recordings of one adult (∼3 years) male budgerigar unfamiliar to the subjects here (one from Group C from [26], see recording details). Each stimulus was thereby composed of the same set of five vocal units, with the only difference between conditions being how they were arranged in time. Using a custom MATLAB script, we generated five rhythmic patterns. Each pattern consisted of five parallel tracks, one for each of the five different vocal units. To achieve this, we created a matrix representing two 4/4 bars, with possible sound onsets at each 16th note. Each pattern consisted of 20−26 events, with at least four events per track. One of the tracks had events on each quarter‐note beat to make sure there was a sound event on each beat in the final stimuli. This way, stimuli could be made rhythmic, such that inter‐onset intervals (IOIs) of subsequent vocal units were related by ratios of 1:1, 1:2, or 1:4 (see Figure 1). All of these rhythmic ratios were present in all stimuli; thus, a preference for specific rhythm ratios could not be tested with the current study design. The duration of each bar was approximately 1.5 s (160 bpm), resulting in about 40 repeating bars per 60‐s stimulus. Thus, fully rhythmic stimuli comprised both regularity and repetition. We then assigned all combinations of the budgerigar vocal units to the five tracks, resulting in 120 stimuli per rhythmic pattern (thus, 600 stimuli per condition). Stimuli were root mean square amplitude normalized and ramped (0.005 s fade in and out). Arrhythmic stimuli were created by introducing random jittering (within a ±200 ms window around the onset) to each of the individual rhythmic stimuli for a total of 600 arrhythmic stimuli. Stimuli were played at approximately 75 dB SPL at the bird's position.

To assess whether fully rhythmic and arrhythmic stimuli differed in acoustic dimensions other than the rhythmic manipulation, we compared stimulus sets for key acoustic features (see Supporting Information, pp. 2–4).

2.1.4. Procedure

Birds received no training for the current study. After starting playback of the test bird's flock recording, the bird was introduced into the testing apparatus and remained there for 2 h. We tested each bird in three experimental sessions. The intersession interval was 1−9 days. For birds to be included in the sample, they had to listen to both sounds at least once across the three sessions, thus allowing them to make informed choices about listening. If a bird did not activate any of the perches in the first two sessions, we discarded these sessions and added a fourth or, if necessary, fifth session before a bird would be removed from the sample. All birds passed the criterion. Assignment of stimulus sets (fully rhythmic, arrhythmic, or silence) to perches was randomized across sessions and individuals. If a bird sat on the perch for at least 300 ms, a randomly selected stimulus from the appropriate set was played. If the bird sat longer than the duration of a stimulus, another randomly chosen stimulus from the same set began immediately. As soon as the bird left the perch, playback ended with a fade out of 50 ms. Across the session, stimuli were selected without replacement until all stimuli were played. During the study period, the same birds also participated in other operant conditioning studies (see above). Thus, birds were not naive to auditory experiments; however, they were naive to the stimuli used here. Additionally, since birds were not rewarded for the current and the previous (Hoeschele and Bowling's) [20] preference study on rhythm perception, we did not expect any biases. We also included “subject” as a random factor in our statistical models.

2.1.5. Statistical Analysis

We first checked whether birds landed on each perch at least once across all sessions to confirm the possibility of informed choice. One female bird did not visit the silent perch; however, that bird generally spent very little time listening to the stimuli (< 10 s in total), probably spending most time at the floor or the walls of the testing apparatus, so we kept it in the sample, taking account for individual differences by including “subject” as a random intercept as well as random slopes across stimulus conditions in the statistical model. We modeled the effect of the birds’ preference for rhythmic versus arrhythmic stimuli versus silence using a linear mixed model with the cumulative duration spent listening to each condition and its interaction with the bird's sex as predictors. We first summed the durations each bird spent listening to each stimulus set across sessions to obtain our response. Because the response distribution was skewed, we then square‐root‐transformed these cumulated durations before fitting the model. All statistical analyses were done in R (version 4.4.0) [27]. We fitted the model using the function lmer (R package lme4; version 1.1‐30) [28]. Dummy coding for sex was done by a custom‐built function. We inspected a histogram and a qq‐plot of residuals and plotted residuals against fitted values but found no strong deviations from the assumptions of normality and homogeneity of residuals [29, 30]. We assessed collinearity by fitting a linear model without the random effect and interaction between fixed effects using the function vif (R package car; version 3.0‐2 [31]), which revealed no collinearity issues. Model stability was assessed by using a leave‐one‐subject‐out (LOSO, jackknife) sensitivity analysis for each random effect (see Supporting Information pp. 6–9 for details. R code generated with AI assistance), revealing the model to be reasonably stable. We obtained 0.95 confidence intervals using the function bootMer (1000 parametric bootstraps, package lme4, version 1.1–30 [28]; see Table S1). For each model, we report the marginal coefficient of determination ($R_{\mathrm{m}}^2$), which reflects the variance explained by the fixed effects alone, and the conditional coefficient of determination ($R_{\mathrm{c}}^2$), which reflects the variance explained by the fixed and random effects combined.

2.2.1. Bayesian Comparison of Hoeschele and Bowling's study and Experiment 1

The previous binary choice study [20] observed a significant effect of sex in rhythmic preferences, with females spending more time listening to fully rhythmic stimuli and males to arrhythmic stimuli. In the current three‐way choice study, we found the same general direction for female preferences. This suggests that there may be an underlying latent preference of females for choosing rhythm, for which both studies provide evidence under different designs, that is, under different data‐generating processes. To estimate evidence for this latent preference, given both studies, we analyzed listening behavior using a hierarchical Bayesian Dirichlet regression, modeling the proportion of total listening time each bird allocated to each stimulus category (R code generated with AI assistance). This approach allowed us to include both experiments despite different choice possibilities (binary choice between fully rhythmic and arrhythmic stimuli in Hoeschele and Bowling [20] vs. three‐way choice including silence in experiment 1), and thus different data‐generating processes.

Because the two experiments differed in the set of available response alternatives and, therefore, represent different data‐generating processes, we modeled each experiment separately with an otherwise identical hierarchical model structure and parameters. We calculated proportions of listening time per condition from the total listening time and included the fixed effects of sex, and random intercepts and slopes initially; however, due to low precision and high standard error, we decided to exclude the correlation between random intercepts and slopes for the final models to avoid overparameterization due to the small sample sizes. Since the Dirichlet distribution is bound between 0 and 1 and thus assumes nonzero values, we added a small constant (0.001) to instances when a bird did not spend any time listening to a stimulus condition.

We fitted all models using Hamiltonian Monte Carlo (R‐package brms [33]) (for model summaries, see Supporting Information, pp. 9–11). Since we observed some divergent transitions, we increased the target average acceptance probability to 0.99, resulting in no divergent transitions. R‐hats were at 1 for all model parameters. Trace plots indicated that chain mixing was good. Posterior predictive checks and scatter plots of observed versus predicted values suggested a reasonable fit for both models. Bayesian R ² were high (Hoeschele and Bowling [20]: R ² = 0.86, 2.5% CI = 0.462, 97.5% CI = 0.993 for both fully rhythmic and arrhythmic proportions; experiment 1: fully rhythmic: R ² = 0.771, 2.5% CI = 0.437, 97.5% CI = 0.961; arrhythmic: R ² = 0.747, 2.5% CI = 0.367, 97.5% CI = 0.960; silence R ² = 0.900, 2.5% CI = 0.643, 97.5% CI = 0.988) and were affected by the small sample size and inclusion of the random intercept. For the data of Hoeschele and Bowling [20], the probability of direction (pd) indicated a chance of 82.1% that female, but not male, budgerigars reliably preferred fully rhythmic (proportion females: 73%, males: 55%) over arrhythmic stimuli. For experiment 1 (three‐choice paradigm), females’ proportion of choosing fully rhythmic (53%) was again higher than males’ (35%), with a 91.1% chance (pd) of a reliable pattern. The differences for choosing arrhythmic stimuli were smaller between sexes and less certain (females: 36%, males: 28%, pd = 74%).

We then computed posterior draws from each stimulus category and calculated mean differences and quantiles at 2.5% and 97.5%. We found that females allocated more listening time to fully rhythmic than arrhythmic stimuli in both Hoeschele and Bowling [20] (posterior mean difference = 0.462, 95% quantiles [−0.317, 0.921], P(Δ > 0) = 0.908) and in the current experiment 1 (mean difference = 0.169, 95% quantiles [−0.114, 0.433], P(Δ > 0) = 0.912). In experiment 1, females also showed a clear preference for fully rhythmic stimuli over silence (mean difference = 0.417, 95% quantiles [0.087, 0.618], P(Δ > 0) > 0.987). In contrast, posterior distributions of male budgerigars for a rhythmic preference were closer to zero in Hoeschele and Bowling [20] (mean difference = −0.095, 95% quantiles [−0.894, 0.809], P(Δ > 0) = 0.420); while males showed a directional preference for rhythm in the current experiment 1 (mean difference = 0.070, 95% quantiles [−0.204, 0.317], P(Δ > 0) = 0.755), indicating no consistent directional latent preference across both studies (see Figure 4).

To assess the evidence that female birds have a latent preference for rhythmic structure across both studies, we compared posterior contrasts for female birds. We found that the probability that data from both studies support a latent female preference for fully rhythmic over arrhythmic stimuli was 0.827. In contrast, the probability of males preferring fully rhythmic over arrhythmic stimuli, given the data of both studies, was 0.315. The probability of males preferring arrhythmic stimuli in both studies was even smaller (0.140).

2.2.2. Summary

We observed that female budgerigars spent more time listening to fully rhythmic than arrhythmic stimuli. When splitting the data and exploring them separately for each sex, females were found to significantly prefer fully rhythmic but not arrhythmic stimuli over silence, while males mostly chose silence, without significant differences in preference. Taking the data of both Hoeschele and Bowling [20] and experiment 1 into account, the two experiments provide converging Bayesian evidence that female budgerigars preferentially attend to rhythmic structure, while male preferences appear inconsistent. Posterior estimates showed a consistent female bias toward fully rhythmic stimuli across both experiments. While uncertainty increased in the three‐choice paradigm, posterior mass remained strongly shifted toward fully rhythmic preference relative to both arrhythmic and silence. Our three‐place preference paradigm, therefore, suggests that the findings with Hoeschele and Bowling's [20] two‐place preference paradigm were likely driven by female preference.

Because we could not discriminate whether female birds’ preference for fully rhythmic stimuli was due to these stimuli being repeated or because they had an underlying regular beat, we decided to run a second study, testing a larger sample of female birds only with stimuli that pitted regularity against repetition. Given that some budgerigars in ten Cate et al. [34] appeared to distinguish regular from irregular sequences (when local features such as absolute time intervals between sound events were kept constant), as well as the results from [20], we hypothesized that regular stimuli would be preferred over silence. However, we did not have a specific hypothesis about whether regularity or repetition would be preferred.

3.1.1. Subjects

Twelve female birds from the same aviary groups as in experiment 1 participated in experiment 2. Five of these participated in experiment 1.

3.1.2. Apparatus

The same apparatus as in experiment 1 was used.

3.1.3. Stimuli

We created two new stimulus sets (see Figure 5) based on the same set of vocalization units from unfamiliar budgerigars used in experiment 1. “Regular only” stimuli were created in the same way as the fully rhythmic stimuli in experiment 1, but with a new pattern introduced every other bar rather than repeating the same two bar sequence (with 35−45 bars within each 1‐min stimulus, depending on the overall tempo), thus avoiding pattern repetition. For “repeating only” stimuli, the same two‐bar pattern was repeated, but with the onset times jittered such that the overall timing did not conform to a regular metrical grid. We used the same number of events (12 per bar) in both stimulus types.

In ten Cate et al. [34], budgerigars trained to discriminate regular from irregular patterns failed to generalize to novel stimuli that were changed in tempo. Because different tempos might be perceived differently and because effects confined to just one tempo may indicate preferences for absolute time intervals rather than generally for regularity versus repetition, each stimulus set here contained four 1‐min stimuli at 140, 150, 160, 170, and 180 bpm (= 20 stimuli per stimulus set). The duration of single bars was between 1.3 and 1.7 s, depending on the bpm. We again assessed acoustic differences between stimulus sets using the same analysis procedure and features as in experiment 1. We did not obtain any meaningful difference (probability outside ROPE > 0.95) for any acoustic feature (see Table S4).

3.1.4. Procedure

The procedure was the same as in experiment 1. Birds were tested in three experimental sessions, each with a break of 1−7 days between sessions. The same birds also participated in operant conditioning studies between sessions. As in experiment 1, the assignment of stimulus sets to perches was randomized across sessions and individuals.

3.1.5. Statistical Analysis

As in experiment 1, we used the cumulative perch duration per condition for each bird as the response variable, again square‐root‐transformed. All birds met the criterion of listening to each sound type (regular‐only, repeating‐only, silence) at least once, such that we did not have to exclude any bird's data. We used the same model parameters as in experiment 1, except without the factor sex and its interaction with sound, because we only tested female birds. Assumptions of normality and homogeneity of residuals were met per inspection of the distribution of residuals and residuals against fitted values. Stability analysis was done as in experiment 1, revealing a substantial influence of one bird (Melanzani, see Supporting Information pp. 12–14). We, therefore, ran a full model including this bird and a reduced model devoid of random slopes without Melanzani and report both results. We also analyzed post hoc whether those individuals who had participated in experiment 1 differed significantly from those who had not. This was not the case, suggesting that there was no effect of experience on preference choice (see Supporting Information, p. 1). The two stimulus sets did not differ across acoustic metrics (see Supporting Information p. 5).

3.2.1. Summary

We found that female budgerigars significantly preferred listening to either repeating‐only or regular‐only stimuli as compared to silence but did not significantly prefer either stimulus type over the other. This suggests that both repetition and regularity are equally interesting for female budgerigars. Comparing the results to experiment 1, female birds spent a similar amount of time with repeating‐only and regular‐only stimuli as they did with the rhythmic stimuli comprising both components in experiment 1. We then decided to conduct the study in humans as well so that we could have a parallel cross‐species comparison. Both regularity and repetition are typical of music across cultures [4, 5]. Repeated rhythmic patterns allow for predictability, based on which surprising rhythmic deviations can occur [35]. Humans also have a proclivity to infer hierarchical meter from musical sequences, a process foundational for the perception of groove and for dance [7]. In turn, repetition without variation might lead to boredom due to lack of surprisal [36]. These observations suggest that humans gain more aesthetic appreciation from regularity over repetition, and potentially even more from the combination of both regularity and repetition.

4.1.1. Subjects

We tested 38 human participants aged between 18 and 43 years (9 males, 29 females, see Supporting Information p. 24). All participants gave informed consent to participate and were financially compensated with 5€ for every 30 min of participation duration (total duration ∼1 h). Participants were fully debriefed after the task and informed that they could object to the use of their data in the study at any time before publishing.

Experiment 3 was conducted in line with the Declaration of Helsinki and only employed methods that are covered by the standard ARI laboratory methodology that was approved by the Commission for Science Ethics of the Austrian Academy of Sciences on April 23, 2024.

4.1.2. Setup

Participants were tested in an anechoic chamber of the Acoustics Research Institute (ARI) of the Austrian Academy of Sciences (Vienna). A designated circular test area within the anechoic room was demarcated with red tape on its periphery to prevent participants from interfering with setups for other studies in this room. Within this circumscribed space, three distinct locations were equidistantly placed with a distance of 120 cm from each other and visually marked by small black‐and‐white checkered mats with drawings of footprints on them to encourage participant interaction. A FE108 speaker (same type as was used in the bird studies) was located on a tripod in the center of the test area, about 70 cm from each mat (see Figure S20). Stepping on a mat for more than 300 ms triggered stimulus playback using an infrared beam interruption system, just like in the bird studies [1]. This was done using a custom‐written Python script and an Arduino Uno REV 3 chip; BCMI. Sounds were played from an HP Pro3500 Series computer (EXPBLAU‐ALT‐Intel Core i5‐3470 CPU). Similar to experiments 1 and 2, in each session of experiment 3, two mats triggered sounds from two different stimulus sets, while the third remained silent. We also monitored participants’ actions during experimental sessions using a preinstalled video surveillance system.¹

4.1.3. Stimuli

We created four different sets of stimuli covering the same conditions used in both experiments 1 and 2—fully rhythmic, regular only, repeating only, and arrhythmic—but replacing budgerigar vocalizations with samples of five instruments from Logic Pro (version 9; Apple Inc.): shaker, clave, cowbell, djembe, and kick drum. Stimuli had a tempo of either 130, 140, 150, 160, or 170 bpm (4 stimuli per tempo, thus 20 stimuli per condition, resulting in 80 stimuli in total), 10 bpm slower than for budgerigars, respectively (see Figure 8). Stimuli were played at approximately 70 dB SPL at the participant's position. Note that because only two stimuli were played in each session (along with a silent choice), humans did both session types from experiments 1 and 2, as well as additional stimulus comparisons.

4.1.4. Procedure

Participants were given written instructions in English explaining that their task was to explore a room in six 5‐min sessions. They were instructed to explore freely within an area around the speaker marked off with floor tape, with the added guidance that they could step on the three distinctly marked spots at their discretion. This guidance was designed to enhance comparability between the human and bird paradigms by encouraging participants to trigger the sounds (budgerigars naturally land on perches and rarely sit on the floor, the latter being comparable to stepping outside of the sound‐triggering locations here). Given that this study included four stimulus conditions (representing those used in both experiments 1 and 2), a Latin square design was used to balance the allocation of stimulus types across participants. This way, each participant could listen to each possible pairwise combination of stimulus types across their six sessions, while both the order of pairwise combinations as well as their assignment to the trigger locations on the floor differed between participants. Participants were allowed to take a break between sessions (max of 5 min) or continue immediately at their discretion. After the experiment, participants were asked to relisten to random sample stimuli from each stimulus set and rate them on 7‐point Likert scales (from “not at all” to “very”) for three questions: (1) How beautiful did you find the audio file?; (2) How interesting did you find the audio file?; and (3) How meaningful did you find the audio file?. Subsequently, participants filled out a questionnaire about their demographics (age, gender, etc.), and musical training (see Table S21 and Figure S19).

4.1.5. Statistical Analysis

A linear mixed model was used to model participant‐level cumulative listening durations across sessions, as a function of stimulus condition in interaction with self‐reported years of musical training (max_instrument_years; when a participant learned several instruments, the longest training experience was taken) and stimulus condition in interaction with gender. We expected no further differences between genders with the same amount of musical training, thus no interaction of gender with musical training. Although the previous study by Hoeschele and Bowling [20] did not find an effect for gender, we included it for a more complete replication and comparison. We also included a random intercept for “subject” as well as random slopes. Our post‐experimental questionnaire showed that most participants had some amount of musical training (79% of 29 females, 67% of 9 males), next to a large skew in gender. We nevertheless included the interaction terms, keeping in mind a potential sampling bias toward females with musical training in our interpretation.

Model diagnostics revealed residuals to be homogeneous and approximately normally distributed, so no data transformation was necessary. We did not find any collinearity issues. The comparison of the initial full model with a null model containing only random intercept and slopes was highly significant (χ ² = 142.514, df = 14, p < 0.001, $R_{\mathrm{m}}^2$ = 0.484, $R_{\mathrm{c}}^2$ = 0.606), providing omnibus protection against type I errors [37]. Due to stability issues, we simplified the initial model (see Supporting Information pp. 14–19). We report a final model, pruned of nonsignificant interactions with only main effects and the random intercept (cumdur ∼ sound + gender + max_instrument_years + (1 | subject)), which was statistically indistinguishable from the initial model.

4.2.1. Summary

We observed that humans preferred listening to fully rhythmic stimuli (combination of regular and repeating) and regular‐only stimuli over repeating‐only and arrhythmic stimuli. A similar pattern existed in the questionnaire responses regarding interest, aesthetic appreciation, and meaningfulness, with a significant preference for fully rhythmic stimuli regarding the latter two questions. This indicates large agreement between behavioral and self‐report measures of preference. Musical training did not exert a significant effect on listening preference, suggesting a low influence on human rhythm preference, which could be addressed in future studies.