Predictive processes shape individual musical preferences

Significance

Using a novel decision-making task, we show that musical pleasure relies on a delicate balance between predictability and uncertainty, consistent with learning theories. In simple terms, music that is not overly expected nor too chaotic is most enjoyable—but the ideal mix of predictability depends on how much the melody keeps you guessing. Very predictable tunes can be delightful with small twists, while a melody full of surprises may need bigger unexpected moments to hit the sweet spot. Computational models incorporating this balance accurately predicted the types of music people like and the pleasure they derive from real compositions. These results reveal fundamental mechanisms driving musical pleasure and offer valuable insights for the music industry and music-based therapies.

Abstract

Current models suggest that musical pleasure is tied to the intrinsic reward of learning, as it relies on predictive processes that challenge our minds. According to predictive coding, optimal learning, which maximizes epistemic value, depends on balancing predictability and uncertainty, implying that musical pleasure should also reflect this equilibrium. We tested this idea in two independent large samples using a novel decision-making paradigm, where participants indicated preferences for melodies varying in surprise and entropy. Consistent with prior research, we found an inverted U-shaped relationship between predictability and preference. Moreover, our results revealed an interaction between predictability and entropy, with smaller surprises preferred in low-entropy melodies and larger surprises favored in high-entropy music, consistent with predictive coding principles. Computational models incorporating this interaction predicted individuals’ genre preferences and pleasure responses to real compositions, highlighting its applicability to real-world music experiences. These findings advance our understanding of the cognitive mechanisms driving music preferences and the role of predictive processes in affective responses.

The ability to drive pleasure through music is one of the most fascinating aspects of human nature. According to a wealth of theoretical models from diverse disciplines (musicology, psychology, philosophy, and neuroscience), music’s emotional impact and enjoyment predominantly arise from the expectations it engenders through its regular patterns and, mainly, when there is a good balance between predictability and surprise (1–4). For instance, syncopations (i.e., musical accents falling outside the musical pulse) lead to a drive to move and increase feelings of pleasure (5). Also, bodily reactions experienced at peaks of pleasure, such as “chills” or “goosebumps,” often appear following unexpected harmonies or subtle changes of loudness (6–10).

Surprises are pivotal in facilitating learning, prompting individuals to reevaluate their existing knowledge and adjust their behavior accordingly (11–13). Indeed, they are a fundamental feature of most learning theories and models, from reinforcement learning to predictive coding (13–16). Under this context, a very appealing hypothesis is that music-induced pleasure results from an intrinsic reward for learning, representing an epistemic pursuit akin to curiosity (17, 18). Acquiring new information is essential for understanding one’s environment and facilitating individuals’ survival, particularly in ever-changing environments. Thus, it seems plausible that because of its biological relevance, the brain may intrinsically reward us when learning occurs and new information is acquired (19), fostering our curiosity and motivation to seek learning challenges. In this regard, the structural and temporal patterns inherent in music make it an ideal stimulus to exploit these learning-related reward responses. Essentially, music may represent a predictive/learning game that our brain loves to play (3, 18, 20).

Consistent with this idea, research from a wide variety of empirical domains has converged on the conclusion that an intermediate amount of predictability may optimize learning (21–24) and, therefore, maximize pleasure. Entirely predictable events prevent learning as they provide no new information, but surprises that are unforeseeable and appear random also hinder learning as they are hard to interpret. In this vein, a manageable challenge (a degree of predictability that is not too high, not too low) may maximize liking compared to low and high levels of predictability in which learning is not achieved (25) (the so-called Wundt effect). The Wundt effect has been shown in studies modulating either harmonic (26, 27), rhythmic (5, 28–30), or melodic complexity (17, 31, 32).

This predictability sweet spot (the degree of predictability that maximizes music pleasure) is likely to be influenced by the uncertainty of the musical composition. According to predictive coding models, not all surprises are equally relevant to guide learning, critically depending on how precise our expectations are. Indeed, surprise signals update prior beliefs in a manner that is weighted by their associated precision (13). In stable scenarios, where predictions are very precise, subtle changes can be highly informative and trigger model updating. Conversely, in more uncertain contexts, where predictions are less precise, surprises are attenuated since they are the norm and larger surprises are necessary to drive learning (33). Thus, uncertainty and the precision of predictions modulates the degree of surprise that optimizes learning. If musical pleasure is tied to epistemic value, one would expect smaller surprises to be preferred in melodies with low uncertainty and larger surprises preferred in melodies with high uncertainty. This pattern is indeed observed in real-world melodies, where uncertainty and predictability are highly correlated (17). Surprisingly, previous studies implementing a computational model of auditory expectation [Information Dynamics Of Music, IDyOM (34)] that captures the degree of surprise (information content) and uncertainty (entropy) in a given composition have shown the opposite pattern: Melodies and chords were found more pleasant if uncertainty was low and surprise was high, or vice versa (17, 27). However, these studies have typically implemented unequal sampling across the spectrum of predictability and uncertainty with a disparity in the number of stimuli representing the different levels of the predictability and uncertainty continuum. Furthermore, an often overlooked complication arises from the inherent correlation between predictability and uncertainty, a relationship that may obscure the independent contributions of these factors. Hence, these limitations underscore the necessity for more nuanced experimental designs by i) increasing the diversity of stimuli, ii) ensuring balanced sampling across the continuum of predictability and uncertainty, and iii) ensuring that the measures of predictability and uncertainty are uncorrelated, thereby disentangling their unique contributions to musical pleasure.

In addition, the predictability sweet spot may differ significantly across individuals. Some people might prefer relatively straightforward, predictable music, while others might thrive on the thrill of more complex and less predictable compositions. For instance, in Western music, some people may enjoy music characterized by repetitive melodic or rhythmic patterns (pop, techno, hip hop, etc.), while others may prefer musical genres involving improvisation and surprise (jazz, experimental music, among others). This variability in musical taste could reflect individual differences in the degree of music complexity that represents a manageable challenge (in terms of both predictability and uncertainty) so that individuals may seek music with a degree of complexity that aligns with their expectations and predictive skills. However, previous studies offered poor or unreliable measures of the Wundt effect at the individual level, likely due to i) the small number of complexity categories (generally three: low, medium, and high complexity/predictability), ii) the relatively small number of stimuli, and iii) relying on self-reporting ratings without clear references.

Finally, while many studies have posited an inverted U-shaped relationship between musical predictability and preference, no study to date has investigated whether this relationship can predict individual preferences for musical stimuli beyond those used to establish the effect. The question thus arises: How applicable is this inverted U-shaped effect in relation to the actual preferences people hold for music in their everyday lives? Is it an artificial construct limited to laboratory settings, or does it translate into tangible, real-world preferences? To do so, it is critical to develop a standardized procedure to assess individual differences in the Wundt effect to test whether the inverted U-shaped hypothesis predicts real-life individuals’ musical preferences.

Here, we aim to address all these limitations by developing a novel decision-making paradigm to assess the contribution of surprise and uncertainty in individuals’ musical preferences (the MUSIc COmplexity Sensitivity task, MUSICOS). By combining artificially generated music (applying Google’s machine-learning tools) and computational models of auditory expectation (IDyOM), a large variety of melodies were classified into 16 categories of music complexity as a function of both predictability (operationalized as information content, IC) and uncertainty (entropy). Both measures were fully uncorrelated by design and evenly sampled, allowing us to disentangle their specific contribution to music preferences. During the task, participants had to indicate their preference between pairs of melodies with different degrees of complexity (in terms of IC and entropy). By implementing computational models of participants’ choice, we aimed to estimate the complexity sweet spot for each individual as a function of both predictability and uncertainty and explore its relation to individual differences in personality, music-related measures, and musical genre preferences. We also aimed to investigate its potential to generate predictions about individuals’ reports of pleasure to different musical compositions (Beethoven sonatas) varying in their complexity. Consistent with previous findings, we hypothesized an inverted U-shaped relationship between predictability and preference. If music-induced pleasure is maximal when learning is optimized, we would also expect this U-shaped relationship to be modulated by uncertainty, following predictive coding principles. Finally, if performance in MUSICOS is transferable to real-world contexts, we expect the individuals’ complexity sweet spot identified by MUSICOS to predict their music taste and subjective reports of pleasure in response to actual musical compositions.

Results

Participants’ Choices Are Influenced by an Interaction Between IC and Entropy.

Using linear mixed modeling (as described in the Materials and Methods section), we investigated the influence of linear and quadratic effects of information content and entropy, as well as their interaction, in participants’ choices. Results revealed a significant main effect for both linear [β = −0.25, SE = 0.044, χ² (1) = 31.57, P < 0.001] and quadratic [β = −0.25, SE = 0.031, χ² (1) = 64.45, P < 0.001] effects of IC. A significant negative quadratic IC effect indicated that there was a significant Wundt effect between participants’ choices and IC. A significant negative linear IC effect indicated that the inverted U-shaped relationship was shifted toward more predictable melodies along the IC continuum. Additionally, a significant main effect of entropy was also found significant [β = −0.12, SE = 0.043, χ² (1) = 7.80, P < 0.001], indicating a preference for melodies with lower over those with higher entropy. No significant effect was found for the quadratic effect of entropy on choice behavior [χ² (1) = 0.49, P = 0.48]. In addition, the interaction between the linear effects of IC and Entropy was significant [β = 0.10, SE = 0.028, χ² (1) = 13.05, P < 0.001]. As shown in Fig. 1A, the IC Wundt curve was shifted toward more surprise as entropy increased. In other words, in melodies with lower entropy the optimal level of IC was lower (smaller surprises were preferred), while in melodies with high entropy, this optimal level shifted toward higher surprises. Interestingly, results showed striking individual differences in the preference for predictability (Fig. 1B).

Fig. 1.

Individuals’ preferences followed an inverted U-shaped relationship with IC which sifted depending on the level of entropy (A) Inverted U-shaped relationship between participants’ choices and IC as a function of Entropy (B) Example of three participants with low (Top), medium (Middle), and high (Bottom) sensitivity to complexity. Note that in all three cases, the inverted U-shape is shifted toward higher IC in high entropy music. (C) Model predictions for the three participants represented in B, with PreSS values of 1.07 (Top), 1.58 (Middle), and 1.96 (Bottom).

A Model of Preference Including an Entropy-Weighted Sweet Spot Explains Participants’ Choices.

To assess the contribution of IC and Entropy on participants’ preference more precisely, three computational approaches were formalized. The overarching assumption of these models is that there exists an optimal level of predictability (the predictability sweet spot, PreSS) that is most preferred by individuals, with deviations from this optimal point, either toward more or less predictability, resulting in decreased preference. The models vary, however, in how they conceptualize the influence of IC and entropy on this sweet spot. The Information Content Preference Model (ICPM) assumes that the melodies’ value is solely influenced by IC and independent of entropy. By contrast, the Entropy-weighted Information Content Preference Model (EwICPM) relies on stimuli entropy-weighted IC. Finally, the Entropy-weighted Sweet Spot Preference Model (EwSSPM) posits that the sweet spot itself, not stimuli IC, is entropy-weighted. The three models, with their variants, estimate individuals’ predictability sweet spot (PreSS). Model comparison indicated that the EwSSPM had the highest chance of being the most common model in the population (with an expected frequency of 0.97 and an exceedance probability, or posterior probability, that it was the more common model of 0.999). As shown in Fig. 1C, the model correctly reproduced individual differences for complexity preference.

Jazz Lovers Showed a Higher PreSS.

Next, the relationship between individuals’ PreSS (estimated by EwSSPM) and i) demographic factors (genre, age, education, musical training), ii) personality measures (extraversion, agreeableness, emotional stability, consciousness, openness, and musical reward sensitivity), and iii) preference for 11 musical genres were explored using linear robust regression analysis. The only predictor that reached significance was subjective reports of preference for Jazz music [β = 0.12, F (1,142) = 10.02, P_BONF = 0.02]. Individuals with a higher preference for Jazz music had higher PreSS, that is, a greater preference for surprises (Fig. 2A). The effect remained significant even when including subjective reports of exposure to Jazz music [Jazz exposure: F (1,141) < 0.01, P = 0.96, Jazz preference: F (1,141) = 4.77, P = 0.03].

Fig. 2.

Individuals Predictability Sweet Spot (PreSS) predicts individuals’ preferences and subjective reports of pleasure. (A) Scatter plot representing the relationship between individuals’ PreSS and preference for Jazz music. (B) Scatter plot representing the relationship between individuals’ subjective reports of pleasure to Beethoven’s Sonatas and the pleasure predicted by the model. (C) Model predictions of preference for each of the Sonata for individuals with Low and High Press. (D) Average subjective reports of pleasure for each Beethoven Sonatas for individuals with Low and High PreSS.

The EwSS Preference Model Predicts Subjective Reports of Pleasure to Real Music.

The utility of EwSSPM in predicting subjective reports of pleasure of real compositions was also explored. Individuals listened to excerpts of four Sonatas by Beethoven, each with varying degrees of IC and Entropy. The EwSSPM was implemented to generate predictions of value for each Sonata and for each individual, given each Sonata’s IC and entropy and each individual’s PreSS. Linear mixed modeling was used to study the relationship between EwSSPM predictions and subjective reports of pleasure. The analysis indicated a significant positive relationship between subjective and predicted reports of pleasure [β = 0.22, SE = 0.077, χ² (1) = 8.07, P = 0.004], indicating that the model’s predictions applied to real music (Fig. 2B).

The EwSS Preference Model Predicts Individual Differences in Musical Preferences.

In addition, a second analysis was conducted to assess the relationship between participants’ preferences for each Sonata and individuals’ PreSS. Individuals with higher PreSS were expected to enjoy Sonatas Nos 29 and 32 more since they were more complex, while individuals with lower PreSS were expected to enjoy the simpler Sonatas Nos 19 and 20 more. Fig. 3C shows the model’s predictions for individuals with high and low PreSS (median split for illustration purposes). Following the model predictions, the analysis revealed a significant interaction between Sonatas and PreSS [χ² (3) = 21.33, P < 0.001]. Fig. 3D shows the actual reports of pleasure of individuals with high and low PreSS, mimicking the models’ predictions.

Fig. 3.

The main findings were replicated in a second sample, demonstrating consistency with the initial results. (A) Inverted U-shaped relationship between participants’ choices and IC as a function of Entropy. (B) Scatter plot representing the relationship between individuals’ PreSS and preference for Jazz music. (C) Scatter plot representing the relationship between individuals’ subjective reports of pleasure to Beethoven’s Sonatas and the pleasure predicted by the model. (D) Average subjective reports of pleasure for each Beethoven Sonatas for individuals with Low and High PreSS.

PreSS as Measured by MUSICOS and the EwSS Preference Model Shows Moderate Reliability.

Finally, the test–retest reliability of PreSS was assessed using interclass correlation. A moderate degree of reliability was found between PreSS measurements. The average measure ICC was 0.673 with a 95% CI from 0.485 to 0.793 [F (75,76) = 3.06, P < 0.001].

The Main Findings Were Corroborated in a Second Sample.

The main results of Study 1 were replicated in a second sample of participants. First, linear mixed-effects models also revealed significant main effects for i) both linear [β = −0.24, SE = 0.035, χ² (1) = 45.56, P < 0.001] and quadratic [β = −0.27, SE = 0.031, χ² (1) = 73.49, P < 0.001] effects of IC, ii) entropy [β = −0.15, SE = 0.043, χ² (1) = 12.90, P < 0.001], and iii) the interaction between IC and Entropy [β = 0.12, SE = 0.028, χ² (1) = 18.64, P < 0.001] (Fig. 3A).

The EwSSPM also emerged as the best model to account for participants’ choice behavior (exceedance probability of 0.999). As in the first sample, individuals with a higher preference for Jazz music showed higher PreSS [β = 0.05, F (1,260) = 5.24, P = 0.02, Fig. 3B]. Finally, the model accurately predicted i) participants subjective reports of pleasure to Beethoven’s Sonatas [β = 0.14, SE = 0.070, χ² (1) = 3.95, P = 0.047, Fig. 3C], and ii) individual differences in preference for each Sonata [Sonatas*PreSS interaction χ² (3) = 10.24, P = 0.017, Fig. 3D].

Discussion

The current study aimed to investigate the contribution of surprise and uncertainty in musical pleasure and its applicability to predict individual differences in musical preferences and pleasure. To do so, a novel decision-making task combining artificially generated music and computational models of auditory expectations was implemented in a way that both IC and entropy were fully uncorrelated by design and equally sampled along their continuum. By using this approach, the current study validated the inverted U-shaped relationship between predictability and preference found in previous studies (5, 17, 24, 26, 28, 29, 35). Individuals preferred an optimal degree of predictability, and beyond and above that optimal level melodies became less preferred. However, the findings also reveal a robust interaction between musical predictability and entropy in shaping individuals’ preferences. This interaction suggests that although listeners have a preference for melodies with an optimal balance between predictability and surprise, this predictability sweet spot is shifted toward more or less surprise depending on the entropy of the musical context. In particular, melodies with smaller surprises were found more pleasant if uncertainty was low too, and larger surprises were preferred in more entropic music. The superior performance of the EwSSPM further indicates that the predictability sweet spot is dynamically calibrated as a function of entropy. The EwSSPM did not only provide a better fit for participants’ choices in MUSICOS but it also accounted for both individuals’ musical preferences and listeners’ enjoyment of real musical compositions, demonstrating cross-stimulus generalization and the model’s applicability to real-world music listening experiences. Overall, these findings support the idea that musical pleasure relies on an intrinsic reward for learning—a process that operates in line with predictive coding principles.

The directionality of the interaction between Entropy and IC is contrary to what previous studies implementing IDyOM have described (17, 27). In particular, these studies found that individuals preferred a higher degree of surprise when uncertainty was low, and a lower degree of surprise when uncertainty was high. Their findings supported a dual process of musical enjoyment, where pleasure is derived either from the resolution of uncertainty in complex musical pieces or from the experience of surprise in simpler ones. Thus, while these studies posit that entropy dictates a preference for resolution versus surprise, our findings reveal that listener preferences shift along a continuum of entropy-weighted predictability. This inconsistency could be driven by the strong correlation between IC and Entropy and unequal sampling of the IC and Entropy continuum in previous studies that could lead to significant biases in the interpretation of how these factors influence listener preferences. Indeed, in both studies, preference was found for those quadrats resulting from the IC and Entropy interaction that were less sampled (high entropy with low IC and high IC with low entropy) (17, 27). The preference for these scarcely represented quadrants could be a result of the statistical models used to analyze the data, which may inaccurately predict listener responses in undersampled areas of the IC and entropy space. Moreover, the negative interaction between IC and Entropy found in those studies seems to be at odds with the positive correlation between IC and Entropy generally found in real compositions¹⁷ (which was also found in the initial sample of artificially generated melodies). If the preference for surprise is lower as entropy increases, one would expect real composition to match this scenario. The use of artificially generated melodies in the current study allowed for the generation of a vast and diverse dataset of musical stimuli and for exploring parts of the IC x Entropy space that may be underrepresented in natural music, thereby yielding a more precise assessment of how these two dimensions interact to shape musical enjoyment.

The adjustment of preference for predictability as a function of entropy resonates with the predictive coding framework (13). At the core of this theory is the concept that the brain operates as a prediction machine, aiming to minimize the difference or error between its predictions and the actual sensory input. When there is a mismatch between predicted and actual sensory input, the brain updates its internal models to better predict future sensory input, thereby learning from the discrepancy. A key idea in predictive coding is that prediction errors are weighted by their expected precision or uncertainty (36). Precision, thus, modulates to what extent prediction errors in a particular context convey “newsworthy” information to drive belief updating. When the environment is relatively predictable and entropy is low (and precision high), even minor surprises can provide significant learning opportunities (33). In contrast, in high-entropy environments where unpredictability is the norm, the brain’s predictive models are constantly challenged by a barrage of surprises. In such contexts, small surprises may not provide enough new information to significantly alter or improve the predictive models. They are effectively lost in the noise. Instead, larger surprises—those that stand out significantly from the ongoing uncertainty—are more likely to make a meaningful impact on the brain’s predictions (33). The importance of precision for predictive processing has been demonstrated in studies of auditory perception measuring the mismatch negativity (MMN) in oddball paradigms. The MMN is an event-related potential (ERP) component that reflects the brain’s automatic response to deviations from auditory expectations, serving as a marker of auditory prediction errors (37). Studies have shown that the amplitude of the MMN to deviants is reduced in more uncertain contexts, providing empirical evidence for the attenuation of prediction errors in uncertain contexts (38–42). Our findings demonstrate that the precision-weighting principle of predictive coding extends beyond auditory perception and may also account for musical pleasure, further indicating that music exploits an intrinsic reward for learning. This intrinsic reward is fundamental across species, offering a clear biological advantage by enhancing adaptation to changing environments (19, 43). Humans, however, may have extended this capacity through abstract rule learning and symbolic representation, which may have broadened the range of activities we find pleasurable—including music.

Critically, the degree of predictability that optimizes learning, and therefore maximizes pleasure, varies across participants. These differences could be the source of individuals’ variation in musical tastes. Notably, the results indicate the individuals’ PreSS predicts participants’ preferences for Sonatas from Beethoven with a diverse range of complexity. Individuals with higher PreSS show higher preference for more complex sonatas Nos. 29 and 32 than individuals with lower PreSS. On the contrary, individuals with lower PreSS showed a higher preference for simpler sonatas Nos. 19 and 20. These findings further strengthen the idea that the EwSSP model and the resulting PreSS can generate predictions about individuals’ preferences for real music composition. Moreover, the results suggest a positive relationship between PreSS and a preference for Jazz music. The higher the preference for Jazz music, the higher the degree of predictability that maximizes pleasure.

The higher preference for surprises in Jazz lovers resonates with previous EEG studies investigating music-processing skills in Jazz musicians(44–47). By using a melodic multifeature paradigm to measure MMN to six types of musical feature violations, Vuust and colleagues showed larger MMN amplitude to all six violations in Jazz compared to Classical and Rock musicians, indicating a greater overall sensitivity to music surprises in Jazz musicians. Similar findings were reported by Kliuchko et al. (2019). Notably, they found that MMN amplitude correlates negatively with Jazz preference in nonmusicians (45), indicating that the effects of training and preference are dissociable in the MMN. Our result, however, addresses a different question with a different population and metric, and therefore offers information those studies could not provide. While previous ERPs studies focused on preattentive perception, they did not address hedonic value. PreSS, in contrast, is derived from a forced-choice preference task and directly quantifies the complexity level that individuals find most rewarding. Notably, PreSS also predicted preference for nonjazz material (Beethoven sonatas), indicating that the trait captures a domain-general rather than a genre-specific sensitivity. Moreover, the link between PreSS and Jazz preference in the current study was independent of subjective reports of exposure, and years of musical training did not correlate with PreSS. These results may indicate the existence of a certain predisposition to enjoy music with higher or lower complexity regardless of music training and exposure. Indeed, if musical pleasure relies on predictive processes, individuals should enjoy music with a degree of predictability that aligns with their predictive abilities (48). Individuals with higher auditory predictive skills could be more likely to enjoy music with a higher degree of complexity if properly exposed. Notably, twin studies demonstrate moderate genetic influences on musicality traits (49) and music specialization—including the type of music individuals engage in (50) —which supports the notion that these traits and preferences may be, at least in part, innate. Longitudinal studies with music exposure interventions (e.g. Jazz) in naïve participants could test the idea that music complexity sensitivity (using PreSS as an index) represents a predisposition to enjoy music with more or less complexity. The moderate test–retest reliability of PreSS, while not perfect, suggests that it could serve as a meaningful construct for understanding individual differences in musical preferences development. Musical preference is known to fluctuate based on factors such as listening context, mood, and cultural exposure, which naturally introduce variability in preference-related measures (48, 51) and likely contribute to the moderate reliability observed for PreSS. Thus, although PreSS may not capture all of these variable factors, it may still represent an important tool for understanding the mechanisms underlying the development of musical preferences.

Notably, results were replicated in two distinct large samples from differing Western cultures, which reinforces the robustness and generalizability of our results. The fact that the results held across these varied samples not only attests to the validity of the findings but also suggests that the observed phenomena transcend cultural and demographic boundaries to some extent. However, the generalizability of our results to non-Western contexts is uncertain. To address these limitations, future research should aim to include participants from diverse non-Western societies, which would provide valuable insights into the cross-cultural applicability of our findings and contribute to a more global understanding of music perception and appreciation.

Despite the strengths of our controlled design, our study has limitations regarding the generalizability of its findings. First, to test cross-stimulus generalization of the EwSSP model we used monophonic excerpts from four Beethoven sonatas. This clear, well-structured melodic framework made IDyOM implementation straightforward but may restrict the extension of our results to richer or more stylistically diverse music. Crucially, however, the EwSSP model was initially calibrated with participants’ choices across 240 artificially generated melodic fragments that are not tied to any particular genre, a feature that helps support broader applicability (52). Second, we did not account for the nuanced differences within genres—for instance, the diverse subgenres of jazz. Although our model may still account for preferences within such styles, future studies could explore the relationship between PreSS and pleasure in those unsampled musical contexts. Nevertheless, current findings provide an important milestone to understand the link between individual differences in music complexity sensitivity and individuals’ musical preferences. Future studies should expand on this work by incorporating a broader array of composers, genres (subgenres), and polyphonic pieces, as well as by implementing novel computational models tailored to polyphonic music (53). Finally, similar decision-making tasks could be implemented to characterize individuals’ sensitivity to rhythm and harmonic complexity, thereby providing a more fine-grained characterization of individual differences in music preferences

In summary, current findings, by combining artificially generated music and computational modeling, shed light on the key role of the intrinsic reward for learning in musical pleasure and the underlying individual differences. In addition, it may pave the way for the development of personalized music recommendation systems to introduce listeners to new genres and compositions that align with their complexity sweet spot, thereby expanding their musical experiences beyond their usual preferences and maximizing listeners’ pleasure. Considering the significant impact of musical pleasure on well-being (54), identifying methods to optimize music reward experiences may be also fundamental for promoting individuals’ welfare.

Materials and Methods

Experimental Design Experiment 1.

Participants.

To achieve the primary objective of investigating the relationship between individuals’ complexity sweet spot and participants’ musical preferences, a power analysis was conducted. The power analysis was based on detecting a moderate effect size (r = 0.30) for correlations, using a two-sided test with a power of 0.95 and an alpha level of 0.05. The analysis indicated that a minimum sample size of 134 participants would be required. To account for potential exclusions due to poor performance or musical anhedonia (55, 56), a total of 180 participants were recruited. Participants (63 women, 118 men; mean age = 34.68, sd = 13.16, range = 18 to 69 y old) were recruited via Prolific and were from English-spoken countries (94 from the United Kingdom, 45 from Canada, 19 from the United States, 11 from Ireland, 9 from Australia). Participants were paid 8 pounds for participation. This study was conducted in accordance with the ethical guidelines from the Declaration of Helsinki and was approved by the University of Barcelona ethics board.

Procedure.

The experiment was run online. First, participants reviewed and provided informed consent. Next, participants completed a series of self-report demographic (age, genre, education, years of musical training), and personality measures [the Ten Item Personality Inventory (TIPI) (57) and the Barcelona Music Reward Questionnaire (BMRQ) (58)]. They also provided subjective reports of preference and exposure to 11 musical genres (Classical, Dance, Rap, Soul, Country, Indie, Jazz, Rock, Metal, Trap, and Pop). After completion, the participants performed the MUSICOS (MUSICOS, described below). Individuals with musical anhedonia (BMRQ < 65) were excluded from the analysis (n = 21). Participants were called for a second (n = 79) and third follow-up session (n = 76). In the second session (2 wk later), participants listened and provided ratings of pleasure for short excerpts of monophonic versions of Beethoven’s sonatas Nos 19,20,29 and 32 (see below). The third session was 6 months later than the first session and participants performed MUSICOS again to assess the test–retest reliability of the task.

Stimuli.

Melodies were generated with Generate, a tool from Magenta Studio, which utilizes a Variational Autoencoder trained on a vast dataset of melodies to generate new, unique melodies by combining learned musical characteristics. Generate incorporates a “temperature” parameter that controls the variability of the generated melodies. Higher temperatures result in more variation and unpredictability, while lower temperatures yield more conservative, predictable melodies. To cover a wide range of musical predictability, initially, 160 piano melodies for each temperature value (from 0 to 2 in increments of 0.1) were generated, resulting in a total of 3,360 melodies of 8.5 s duration.

To objectively quantify the statistical predictability of melodies, the unsupervised variable-order Markov model IDyOM (Information Dynamics Of Music (34)) was implemented, using the same configuration detailed by Gold et al., 2019 (17). IDyOM has consistently demonstrated its ability to accurately assess objective indicators of pitch unpredictability/surprise (operationalized as Information Content, IC) and uncertainty (as represented by entropy) in Western listeners (17, 27, 59, 60). On a note-by-note basis, IDyOM reads musical sequences and generates a probability distribution for the following events from which entropy and information content are computed. These computations were the result of a combination of long- and short-term models (LTM and STM respectively). The LTM was pretrained on a large corpus of Western music (185 chorale melodies harmonized by Bach), mimicking the implicit statistical musical knowledge acquired over a lifetime by listeners. The STM, on the other hand, was trained online for each musical piece, simulating the statistical learning from each particular piece. The mean information content and entropy for each stimulus were calculated by averaging the information content and entropy of all notes in that melody. Ninety-five melodies (2.8%) were excluded because they resulted in extremely low or high IC or entropy. The temperature parameter of Generate showed a positive correlation with IDyOM’s IC [r (3265) = 0.46, P < 0.001) but not with Entropy [r (3,265) = −0.03, P = 0.09]. In addition, both IC and entropy showed a positive correlation [r (3265) = 0.62, P < 0.001] similar to other studies (17).

Next, a range of IC (from 4.5 to 10.5) and Entropy (from 4.3 to 5.5) was determined in a way that i) aligned with those used in previous studies, ii) allowed IC and entropy to be orthogonal, and iii) allowed for equal sampling across both dimensions. The IC and Entropy ranges were then divided into eight and two levels, respectively, creating 16 distinct categories or quadrants within the IC and Entropy continuum. From each category, 15 melodies were randomly selected for inclusion in the task, for a total of 240 melodies (Fig. 4A). The correlation between IC and Entropy among selected melodies was r (240) = 0.02 (P = 0.76). All stimuli are publicly available at: https://osf.io/rgp6e/?view_only=c71a0666ddcf477f9989850b7ec3b535

Fig. 4.

(A) Stimulus IC and Entropy distribution. Note that both IC and Entropy were uncorrelated by design and equally sampled across their continuum. Melodies were classified into 16 categories as a function of IC and Entropy. (B) Task schema of one trial. Participants listened to a pair of melodies from different complexity categories and had to decide which one they preferred.

The MUSICOS.

The task consisted of 120 trials where pairs of melodies from different categories (in terms of IC and/or Entropy) were compared using an all-against-all strategy. In each trial, a pair of melodies were presented sequentially, and immediately after, participants had to indicate which one they preferred by pressing the keyboard button 1 (if they preferred the first melody) or 2 (if they preferred the second melody). Participants were forced to respond in less than five seconds (Fig. 4B). A number indicating whether the first or second melody was played (1 or 2, respectively) was presented on the screen while the music was playing. For each trial, a pair of categories was randomly selected with the constraint that each category pair was presented once without repetition. The order of presentation within each pair was randomized. Melodies representing each category were also randomly selected from the pool of melodies within each category, with no repetition. The task also included 5 catch trials to ensure that participants were actively engaged and correctly following instructions. In catch trials, only one melody was played in either the first or second position. In the other position, the corresponding number appeared on the screen, in silence, for 8.5 seconds. In this case, participants were instructed to press the button corresponding to the position at which the melody was played. An a priori>80% accuracy cut-off in catch trials was set to ensure participants were paying attention to the task. Following this rule, seven participants were excluded. Participants who missed 5 or more trials (n = 6) or showed a systematic response bias (choosing either response 1 or 2 for 75% or more of their responses, n = 2) were also excluded (final sample: n = 144). After a practice session, participants proceeded with the full task, which lasted around 40 min.

Music listening.

In a second follow-up session participants listened and reported subjective reports of pleasure to four monophonic excerpts from Beethoven’s sonatas (63 s to 75 s), two from his middle period [Sonatas Nos. 19 (third movement) and 20 (first movement)] and two from his later period [Sonatas Nos. 29 (fourth movement) and 32 (first movement)], representing a spectrum of complexity in his compositional evolution. The later sonatas exhibit a higher degree of complexity than previous periods. IDyOM was then implemented to quantify the degree of complexity of each excerpt. As expected, Sonata Nos. 19 and 20 had lower IC (Sonata 19 = 4.50; Sonata 20 = 4.24) and Entropy (Sonata 19 = 3.55; Sonata 20 = 3.32) than Sonata Nos. 29 and 32 (IC: Sonata 29 = 7.94; Sonata 32 = 5.32, Entropy: Sonata 29 = 3.97; Sonata 32 =3.57). While listening to the music participants were instructed to provide real-time ratings of subjective pleasure by continuously adjusting a slider with the mouse. For each participant, real-time ratings of pleasure were normalized across Sonatas. Average ratings were used for further analysis.

Statistical analysis.

To test the contribution of IC and Entropy in participants’ choice in MUSICOS, we performed linear mixed modeling in R (version 4.0.2) (https://www.npackd.org/p/r/4.0.2) and RStudio (https://www.rstudio.com/) using the lme4 package (61) with participants’ choices as dependent variable. We used linear mixed modeling (LMM) to analyze the data, as it allows for the inclusion of participant-level random effects while accounting for repeated measures within individuals. Given that our predictors were uncorrelated by design, multicollinearity was not a concern, making LMMs an appropriate choice for examining their independent contributions to musical pleasure. Based on previous findings showing the linear and quadratic effects of IC, Entropy, and their interaction in participants’ music liking (17, 27), a model including linear and quadratic effects of IC and Entropy, as well as the interaction between Entropy and the linear and quadratic effects of IC was generated. Stimulus was included as a random intercept to account for variability in participants’ choices that could be attributed to differences in each stimulus. Participants’ random slopes were also included following a three-step strategy. First, the maximal random effects structure was fit, including all within-participants random slopes and their interactions, as recommended to account for all sources of variability and minimize Type I error (62). If the full random structure model did not converge, correlations between random slopes were then removed, thereby simplifying the covariance structure without discarding valuable variance information. Finally, if the resulting model still did not converge, all random slopes accounting for less than 1% of the variance were removed (63). This final step ensures that only meaningful sources of variability are modeled, reducing overfitting. This three-step strategy always resulted in convergence. The effects of the different predictors and their interactions on participants’ choices were assessed through likelihood ratio tests using the car package in R (https://cran.r-project.org/web/packages/car/index.html). These tests were based on Type 3 sums of squares.

Robust regression analysis using the rlm function of the MASS package in R was performed to assess the relationship between the free parameters of the best computational model of preference (see below) and i) demographic variables (genre, age, education, musical training), ii) personality measures (extraversion, agreeableness, emotional stability, consciousness, openness, and musical reward sensitivity), and iii) subjective reports of preference toward eleven musical genres. Bonferroni correction was applied to correct for multiple comparisons within each category (demographics, personality factors, musical preferences). To further explore whether the correlation between the models’ parameters and musical preferences was influenced by musical exposure, a separate robust regression analysis was performed for each musical genre where preference was a significant predictor, including both preference and exposure as independent variables.

To test the relationship between subjective reports of pleasure to Beethoven’s sonatas and the predictions of computational models of musical preferences, mixed-effect models were implemented using the lme function of the lme4 package. Standardized reported pleasure for each sonata was included as the dependent variable. The model’s prediction for each sonata was included as a fixed factor. Stimulus was included as a random intercept to account for variability in participants’ choices that could be attributed to differences in each sonata. Within-participants random slopes for predicted pleasure were also included. In addition, to assess whether the participants’ sweet spot derived from the computational model could account for individual differences in preference for the different sonatas, a second mixed-effect model was implemented including standardized reported pleasure as the dependent variable, and sonata number, PreSS (Predictability Sweet Spot; see below), and their interaction as fixed factors. Stimulus was included as a random intercept.

Finally, interclass correlations (ICC) to assess the test–retest reliability of the parameters estimated by the model were implemented using the ICC function from the IRR package in R.

Computational models of preference.

To further explore the main findings of the linear mixed-effect models, we implemented and compared three preference models on participants’ choices. The three models were grounded in the assumption that there is an optimal degree of predictability—referred to as a “sweet spot”—that individuals prefer. Deviations from this optimal level, either toward more predictable or less predictable information, lead to a decrease in preference. The three models, however, vary in how they conceptualize the influence of IC and entropy on participants’ choices.

The ICPM posits that the optimal predictability sweet spot and therefore, stimulus value, is solely determined by the Information Content (IC), regardless of entropy. Stimulus’ value (V_sx) is calculated as the quadratic deviation of the stimulus IC from the participants’ sweet spot (Predictability Sweet Spot, PreSS) for each stimulus in a pair. The larger the deviation, the lower the value. PreSS is a free parameter estimated for each participant:

$$V_{\mathit{sx}}=-(abs({\mathit{IC}}_{\mathit{sx}}-PreSS){)}^2.$$

The Entropy-weighted EwICPM assumes that the perception of predictability is influenced by entropy (Ent), by entropy-weighting stimulus IC (EwIC).

$${\mathit{EwIC}}_{\mathit{Sx}}={\mathit{IC}}_{\mathit{sx}}/{\mathit{Ent}}_{\mathit{sx}}.$$

Similar to the first model, the stimulus value is then calculated as the quadratic deviation of entropy-weighted IC from the sweet spot:

$$V_{\mathit{sx}}=-(abs(Ew{\mathit{IC}}_{\mathit{sx}}-PreSS){)}^2.$$

The EwSSPM posits that the sweet spot (rather than the stimuli IC as in EwICPM) varies with the entropy of stimuli, by dynamically adjusting the sweet spot as a function of each stimulus entropy. Stimulus value is then calculated as the quadratic deviation of IC from the entropy-weighted sweet spot:

$$V_{\mathit{sx}}=-(abs({\mathit{IC}}_{\mathit{sx}}-PreSS\ast {\mathit{Ent}}_{\mathit{sx}}){)}^2.$$

In all three models, the probability of choosing a stimulus was then modeled using a softmax function, where the temperature parameter (beta) controls how deterministic the choices are.

$$p(s1)=\frac{e^{V_{s1}\ast Beta}}{e^{V_{s1}\ast Beta}+e^{V_{s2}\ast Beta}}.$$

The model’s free parameters were optimized using the fmincon function in Matlab R2021 to search for a parameter set that minimized the discrepancy between the empirical data and the model’s predicted response on each trial, a process that was repeated ten times on each participant with random starting points to avoid local minimal. Model fits were compared using Bayesian model comparison (spm_BMS function in SPM12) (64).

The winning model was also used to generate predictions of participants’ subjective reports of pleasure to Beethoven’s Sonatas by using the corresponding formula to estimate stimulus value using the IC and Entropy of each sonata and the PreSS scores estimated with MUSICOS.

Experimental Design Experiment 2.

Participants.

To test the robustness and generalizability of the initial findings, results from experiment 1 were replicated in a second sample. To refine the sample size estimation, the smallest observed significant correlation coefficient (r = 0.23) was used to determine a more accurate sample size. This updated power analysis indicated that a sample size of 235 participants would be required to detect this effect size with adequate power. Therefore, we employed a second sample to meet this requirement and ensure the robustness of our findings. Two hundred seventy-six students from Barcelona (228 women, 48 men; mean age = 22.69, sd = 6.30, range = 18 to 56 y old) were recruited to further assess the generalizability of our findings across different cultural contexts. Participants were paid 10 euros for participation. This study was conducted in accordance with the ethical guidelines from the Declaration of Helsinki and was approved by the University of Barcelona ethics board. Five participants were identified as musical anhedonics (BMRQ < 65) and were excluded from the analysis. The incidence of musical anhedonia in our second sample was notably lower (~2%) than the first sample (~12%). This discrepancy may be partly explained by the relative homogeneity of the second sample—comprising predominantly young university students—and potential selection biases, as this group may be more inclined to participate in studies they find engaging, thereby attracting fewer individuals with reduced musical reward sensitivity. One participant had <80% accuracy in catch trials and was also excluded. Eight participants showed a clear response bias and were also excluded. The final sample was of two hundred sixty-two participants.

Procedure.

The experiment was run online. However, in contrast to Experiment 1, participants met the researcher via Zoom where the task was explained in detail, and after the meeting, participants had 1 h to complete the experiment. The rest of the procedure was the same as in Experiment 1. Among all participants, 75 participated in an on-site follow-up session in which they listened to and rated Beethoven’s Sonatas as in Experiment 1. During this follow-up session, participants also completed a series of auditory tasks for the purpose of another study.

Statistical Analysis.

The same analyses as in Experiment 1 were implemented.

Data, Materials, and Software Availability

Anonymized Behavioral data and the scripts used have been deposited in OSF (https://osf.io/rgp6e/) (65).