An early study examining the relative importance of different emotional sources in music compared singing with instrumental versions (Stratton and Zalanowski, 1994). It found that, in the absence of singing, listeners were unable to identify the intended emotion of the music. Even when the melody was upbeat, the presence of sad lyrics heightened feelings of depression in listeners. The authors concluded that lyrics play a more significant role than vocals or instrumental elements in shaping listeners' emotional responses. These findings align with research suggesting that lyrics also dominate song memory (Peynircioğlu et al., 1998; Serafine et al., 1984).

However, these findings contrast with those of another study that examined the potential dominance of lyrics in shaping affective responses by comparing lyrics without vocals to instrumental music (Sousou, 1997). In this study, participants read happy or sad lyrics while listening to either happy or sad instrumental music. The participants' mood was influenced by background music rather than lyrical content, suggesting that, in the absence of vocal cues, instrumental music exerted a stronger effect on affective responses than lyrics alone. And in yet another study, when a song in an unknown language was presented in which sad lyrics were accompanied by happy music, happiness ratings were greater than sadness ratings. But when the translation of the sad lyrics was also presented on the screen, neither music nor lyrics dominated and ratings of happiness and sadness were statistically indistinguishable (Mori and Iwanaga, 2013).

Similar conclusions were drawn in a separate study (Ali and Peynircioğlu, 2006), which compared the emotional influence of instrumentals with that of vocals and lyrics. Here, the intended emotion of the instrumental track was found to dominate listeners' affective responses, overriding the emotional content of added vocals and lyrics. Notably, the authors also observed an interaction between the emotion conveyed and the presence of vocals and lyrics: when these were emotionally congruent with the instrumental, they intensified sadness but diminished feelings of happiness. Several studies have reported a distinct processing of happy and sad music, whereby the dominance of vocals and lyrics increases in sad music (Brattico et al., 2011; Vidas et al., 2020). However, this distinction is not found in all cultures (Barradas and Sakka, 2022).

Research on song perception has also explored whether the concurrently presented auditory streams of lyrics and melody are processed independently or as an integrated whole. Neurological evidence points to both shared and specialized pathways for music and language processing.

The view of integrated processing is buttressed by converging evidence from neuroimaging and behavioral studies. These studies report neural overlap in the processing of speech and music (Rogalsky et al., 2011; Schulze et al., 2011), shared neural circuitry (Tallal and Gaab, 2006), overlapping cortical and subcortical networks (Asaridou and McQueen, 2013), and even neural populations selectively responsive to sung music located adjacent to regions specialized for speech and instrumental music (Norman-Haignere et al., 2022). Additional evidence points to overlapping mechanisms for semantic processing in language and music (Calma-Roddin and Drury, 2020; Miranda and Ullman, 2007), as well as between speech and song (Rossi et al., 2020). Consistent with this view, Samson and Zatorre (1991) found that songs are better remembered when accompanied by lyrics.

Other research however supports the independence hypothesis. Bonnel et al. (2001) found no interference between semantic and melodic processing in a divided attention task, implying that lyrics and melody may be processed as distinct perceptual Gestalts (Bregman, 1990). Likewise, Besson et al. (1998) provided evidence for the independence of linguistic and melodic components in songs. A more nuanced perspective is offered by Sammler et al. (2010), who demonstrated interactive processing of lyrics and melodies in the left middle superior temporal sulcus at a prelexical level, but a functional dissociation in more anterior regions. This supports the idea of a posterior–anterior gradient, along which integration and separation vary. Still, the degree of integration or separation might be task-dependent (LaCroix et al., 2015).

On the whole, the existing literature suggests that the contributions of lyrics and music to emotion processing in song are unequal, thereby supporting the independence hypothesis. As outlined earlier, some older studies indicate that lyrics exert a stronger influence on affective responses (Stratton and Zalanowski, 1994; Peynircioğlu et al., 1998; Serafine et al., 1984), while others find that music plays the leading role (Sousou, 1997; Ali and Peynircioğlu, 2006). More recent studies suggest that the relative influence of lyrics and music may depend on the specific emotional content being conveyed (Brattico et al., 2011; Vidas et al., 2020).

However, these studies face methodological limitations. For instance, research using popular and familiar music (Brattico et al., 2011; Vidas et al., 2020) cannot fully control for genre-specific trends toward positive affect in popular music (Interiano et al., 2018), nor for personal memories associated with familiar songs. Such memories can themselves evoke emotional responses, and when triggered by music, they are often associated with positive affect (Cuddy et al., 2017; Jakubowski et al., 2020; Krumhansl and Zupnick, 2013; Parks and Clancy Dollinger, 2014). Moreover, in some studies, the instrumental stimuli were drawn from instrumental pieces (Ali and Peynircioğlu, 2006; Sousou, 1997), which were never intended to accompany vocals or lyrics. This may have introduced a bias in favor of instrumental dominance in affective responses.

Our study seeks to address these limitations through rigorous stimulus validation, while also extending the focus beyond emotional responses to examine broader conceptual processing in song. While studies investigating event-related potentials suggest that instrumental music can evoke concepts beyond emotion, such as “wideness” (Koelsch et al., 2004) or “battle” (Proverbio et al., 2022), to the best of our knowledge, no study has studied the processing of semantic content when presented simultaneously both in musical and linguistic means.

For the conceptual priming experiment, we recruited 50 participants (31 identified as male; as age was assessed in brackets, we report the median age group of 20–40 years here as well as for the participant groups below). 16 participants completed the experiment in a laboratory setting, while the remaining participants were recruited through mailing lists. For the online group, a brief pre-test meeting was conducted via Zoom to establish rapport and explain the procedure.

Before the lexical decision task, participants completed a demographic questionnaire and an integrated informed consent form on the Gorilla™ platform. Participants also answered a number of ad hoc questionnaire items, which aimed to assess subjective bias toward different components of the songs (see Supplementary Table S1). Participants received compensation of 12 euros (or the local currency equivalent) for their participation.

We selected songs by Boris Vian based on the assumption that, even in France and other French-speaking regions, he is predominantly known as a writer rather than as a musician, and even less so as a composer. We restricted our stimulus set to 13 songs for which Boris Vian was both the lyricist and composer, and for which musical scores were available. The arrangements for voice and piano were obtained from Éditions Jacques Canetti. To the best of our knowledge, several of these songs were recorded for the first time by our team, allowing the musicians a high degree of interpretative freedom during the recording process. Participants' familiarity with Vian's musical work was low: Only 10% of all participants reported being familiar with Vian's music (see Supplementary Table S2). This supports the conclusion that our stimuli were unlikely to evoke personal memories or associations, minimizing potential confounds related to prior exposure.

From the pool of 13 songs, excerpts were selected such that approximately half of the selected excerpts featured passages in which the lyrical and musical semantic content differed. In some cases, different lyrics were set to identical musical accompaniments. These selections were made based on the subjective judgment of AK, and subsequently tested through two preliminary studies (detailed below).

In total, 25 excerpts were selected, each averaging 17 seconds in duration. Each excerpt was recorded in three distinct versions:

The M and L versions were used to generate target words and were evaluated in preliminary studies regarding a number of variables (detailed below). The ML versions—combining both linguistic and musical information—served as the prime stimuli in the conceptual priming task.

To derive appropriate target words for use in the priming task, we recruited two additional groups of participants (N = 99, 60 male, median age group: 20–40 years). The subjects were asked to provide free associations in response to auditory excerpts from each of the three stimulus versions. Participants were instructed to list up to four words that spontaneously came to mind while listening to each excerpt. They were encouraged to respond quickly and intuitively, in order to capture immediate conceptual associations and minimize the influence of reflective or strategic thinking. The first group (N = 47) evaluated excerpts from version ML, while the second group (N = 52) listened to excerpts from versions M and L. All participants were recruited online.

Following data cleaning (lower-casing, lemmatization, and normalization), the most frequently mentioned associations were grouped into semantically homogeneous categories based on their conceptual proximity. This categorization was performed by AK using semantic analysis, which examines the meanings of words and their relationships in order to group them into shared semantic fields, and content analysis, which systematically categorizes words based on emergent thematic criteria to interpret patterns of meaning. Despite potential biases, manual annotation remains a standard and accepted practice in studies of this kind (Bolognesi et al., 2017). From these categorized association fields, the most representative terms were selected as target words for the main experiment (see Supplementary Table S3).

We employed a cross-random-effects design using a lexical decision task, in which participants judged whether a given letter string was a real word or a pseudo-word. The primary objective was to compare reaction times (RTs) to real target words originating from excerpt versions M and L, presented in response to the ML recordings which were the primes. RTs were interpreted as an index of conceptual proximity between prime and target—shorter RTs were taken to reflect a stronger semantic association.

We randomized the trial order for each participant. During each trial, a fixation cross appeared at the center of the screen where the target word would subsequently appear. Participants were instructed to fixate on the cross to minimize eye movements and enable faster responses. They were asked to decide as quickly as possible whether the displayed word was a real French word (pressing the “right” arrow key) or a pseudo-word (pressing the “left” arrow key). If a response took longer than 500 ms, a “Answer faster!” prompt appeared.

Following each trial, a burst of white noise was played. This served two purposes: to mask the auditory trace of the preceding excerpt and to function as a pacing mechanism. Participants could only proceed after pressing the space bar, ensuring attentiveness and allowing them to advance at their own pace. They were instructed to keep their right-hand fingers positioned on the “left” and “right” arrow keys throughout the task to enable rapid responses.

The estimated duration of the test was approximately 45 min. However, since participants progressed at their own pace, the actual durations varied. We provided headphones for in-lab participants and encouraged remote participants to use headphones as well.

Our first analysis examined the central question of whether musical or lyrical meanings are accessed more quickly when participants are primed with a song. To assess this, we first log-transformed the RTs in the lexical decision task to reduce positive skew and approximate a normal distribution, in accordance with standard practice in the field (Kousta et al., 2009). Prior to statistical analysis, we removed outliers using the IQR method. Figure 2 presents violin plots summarizing the distribution of log-transformed RTs.

Next, we accounted for a possible influence of the prime properties by controlling for their inner congruency using two separate scores: the Euclidean distance and the cosine similarity between their two constituting components, that is, the distance between versions M and L as recordings in the Euclidean distance score, and as conceptual associations in the cosine similarity score. We further controlled for a possible influence of target word properties, namely, the latent semantic properties (Section 2.3.2) and the lexical properties (Section 2.3.3). We also controlled for a possible influence of the semantic distance between prime and target and a potential interaction between these two effects. In order to account for individual differences between participants, we first explored the correlation between their preferences toward one or the other song component in song listening, as self-assessed in an ad hoc questionnaire, and their implicit bias, as indexed using their RTs. Second, we evaluated the possible influence of individual variability of participants by controlling for the effect of demographic variables.

To assess the effect of stimulus version on affective ratings, we fitted a maximum likelihood mixed-effects model with random intercepts for both participants and items (see Figure 3). The reference condition was set to version ML. For valence ratings, version M was associated with significantly higher ratings compared to version ML (β = 0.54, p = 0.023), Bonferroni-corrected α = 0.025 while version L showed a weaker and non-significant effect (β = −0.32, p = 0.175). For arousal ratings, while the trend remained the same, neither of the comparisons was significant (version M: β = 0.44, p = 0.089, version L: β = −0.23, p = 0.3).

We fitted a linear mixed-effects model using maximum likelihood estimation to examine the effect of target version on RTs, while accounting for random intercepts for participants and primes. A random slope for participants was not included due to convergence issues and model overfitting.

The model demonstrated good fit (AIC = −1, 476.9, BIC = −1, 449.0, log-likelihood = 743.5). Random effects showed a variance of 0.034 for subject intercepts and 0.002 for the primes, while the residual variance was 0.025. The fixed-effect intercept was estimated at 6.456, while the effect of the version L-derived target words (lyrics associated concept) was estimated at 0.0294 (p < 0.001), reflecting a statistically significant increase in RTs relative to version M-derived target words (music associated concept), corresponding to 17 ms (see Table 1).

To validate the contribution of the target version predictor, we also fitted a reduced model without this variable. A likelihood ratio test between the two models confirmed a significant difference (χ(1)2=16.756,p<0.001), underscoring the explanatory power of target version.

The marginal Rm2=0.0036 and conditional Rc2=0.596, calculated using the r.squaredGLMM function from the MuMIn package (Bartoń, 2024), indicated that while the fixed effect accounted for a small portion of the variance, the random effects contributed substantially. Cohen's d, calculated manually, was 0.57, indicating a moderate effect size.

We estimated confidence intervals using the confint function from the lme4 package (Bates et al., 2015). The 95% CI for the intercept (version M-derived target words) ranged from 6.40 to 6.51, reflecting a precise estimate. For version L-derived target words, the 95% CI ranged from 0.015 to 0.043, suggesting a small but reliable positive effect, as the interval does not include zero.

To assess the possible influence of prime properties on the semantic recognition of targets, we calculated its inner semantic congruency score based on each excerpt's position in a two-dimensional space defined by valence and arousal ratings (see Section 2.2.1). The effect of inner congruency was not statistically significant (p = 0.156). Similarly, we controlled for the congruency between M and L versions of the prime excerpts using the cosine similarity scores (Section 2.3.1). Adding this predictor did not yield a significant effect either (p = 0.867).

We further tested a congruency hypothesis from earlier studies, including Ali and Peynircioğlu (2006), which reported that emotionally congruent combinations of lyrics and melody intensified perceived emotion. In our study, we used Euclidean distance in valence–arousal space as a proxy for congruence between excerpt version M and L and examined its effect on excerpt version ML ratings. While Euclidean distance did not predict valence ratings (p = 0.839; R²m = 0.0018), it significantly predicted arousal ratings (p = 0.0304; R²m = 0.188), indicating that greater incongruence between music and lyrics was associated with increased emotional arousal.

To account for the semantic characteristics of the target words, we initially attempted to fit a full mixed-effects model including all latent semantic properties as fixed effects, along with target version and random intercepts for participants and primes. However, due to multicollinearity among the semantic variables, we instead ran separate models for each one, using version M-derived target words as the reference level.

Across all models, the estimated effect of target version changed slightly but remained statistically significant in all but the models controlling for subjective frequency, imagebility and valence (see Table 1 for the estimates for the effect of version in the basic model and the models with additional predictors). To evaluate the impact of these three predictors, we conducted ANOVA model comparisons showing that none of those models significantly improved model fit (see Table 2).

Overall, none of the latent semantic predictors accounted for additional variance in RTs beyond that explained by target version. The RT difference between versions M and L thus remained robust even when controlling for these semantic factors.

To evaluate the impact of the lexical properties of the target words on reaction times, we fitted mixed-effects models with random intercepts for participants and primes. The target version effect remained statistically significant in all models (see Table 1).

Only two models yielded additional significant effects, namely lemma frequency in films and number of syllables. Firstly, the lemma frequency in films had a significant negative effect on RTs (t₍₇₃₎ = −4.515, p < 0.001), indicating faster responses to more frequent words (R²m = 0.014, R²c = 0.59). Model comparisons confirmed that including the frequency in films predictor significantly improved model fit (see Table 2). Type III ANOVA indicated that both frequency in films and version predictors contributed significantly, with the frequency in films exerted a slightly greater impact (F = 20.39 for the frequency in films vs. F = 13.39 for target version), consistent with prior psycholinguistic findings on word frequency.

Secondly, the number of syllables had a significant positive effect on RTs (t₍₂₉₎ = 4.383, p < 0.001), indicating slower responses to words with more syllables. Type III ANOVA revealed that number of syllables accounted for variance comparable to the target version effect (F = 19.20 vs. 19.60). Model comparisons confirmed improved fit with this predictor (see Table 2).

In order to control for a possible influence of the semantic distance between primes, that is, the ML recordings, and target words, that is, concepts derived from M or L recordings, we calculated their Euclidean distance using the valence and arousal scores obtained from pre-testing (for primes see Section 2.2.1, for target words see Section 2.3.2). A paired t-test showed that the difference between the two distances, that is, ML prime to version L-derived target word vs. ML prime to version M-derived target word, was not statistically significant (t₂₄ = 1.82 p = 0.081).

Adding the semantic distance predictor to the model revealed its significant influence on the RTs (t₍₅₂₃₎ = 2.058, p = 0.04). The further apart the affective profile of the target word is from the prime, the longer it takes participants to respond, suggesting that emotional distance between the prime and the target makes processing less efficient. The type III ANOVA showed that the effect of version remained statistically significant even when controlling for the semantic distance (F_{(1, 1, 999.81)} = 19.16, p < 0.001). Given that target words derived from version M recordings tended to be significantly higher rated on valence and arousal than target words derived from version L recordings (see Section 2.3.2), we also checked for a potential interaction between the effect of target version and semantic distance. The interaction was significant as confirmed by type III ANOVA (F_{(1, 1, 679.89)} = 6.13, p = 0.013) suggesting that the effect of semantic distance on RTs differs between version M- and version L-derived target words.

To follow up the significant interaction between target version and semantic distance on RTs, we conducted a simple slopes analysis using estimated marginal trends from the interactions package (Long, 2024). The analysis revealed that semantic distance had a significant positive effect on RTs in version M (b = 0.037, 95% CI [0.0135, 0.0606]) indicating that participants responded faster to the version M-derived target words when they were affectively closer to the primes (ML recordings). In contrast, no such effect was observed for version L-derived target words (b = −0.0009, 95% CI [−0.027, 0.025]) suggesting that emotional distance from the primes did not influence reaction times when the target word was derived from a lyrics only recording (see Figure 4).

Participants' explicit biases were measured using responses from the self-report questionnaire (see Supplementary Table S1). Preference scores were mean-centered, such that positive values indicated a bias toward M and negative values toward L. To assess implicit bias, we computed the difference in participants' log-transformed RTs between version M- and version L-derived target words. This RT difference served as an index of automatic, implicit preference for one component over the other.

A Pearson correlation between explicit and implicit bias measures revealed a weak but positive association (r = 0.269, p = 0.059), suggesting that participants who explicitly preferred lyrics also tended to show slower responses to version L-derived target words, although the relationship was modest (see Figure 5).

Finally, we explored whether demographic variables (sex, age, education level, musical experience, and music liking) could explain additional variance by incorporating them as predictors. None of these variables showed a statistically significant effect, indicating that individual differences of participants did not account for more variance than the main condition effect.