Rhythm is crucial for the perception and production of vocal communication in both music and language. In language, syllable rhythms aid in the segmentation of speech (Cutler and Butterfield, 1992; Dilley and McAuley, 2008), convey the meaning of the speaker through prosodic stress (e.g., sarcasm, Cheang and Pell, 2008), illustrate the presence of a foreign speakers’ accent (Polyanskaya et al., 2017), and support simultaneous acquisition of multiple languages in infancy (Werker and Byers-Heinlein, 2008). In music, rhythm contributes to melodic identity (Jones et al., 1987; Hébert and Peretz, 1997), enables beat perception (Povel and Essens, 1985; Parncutt, 1994), impacts perceived groove in music (Matthews et al., 2019), and provides the structure that allows synchronization with music or other people (Fitch, 2016). Rhythm is clearly an important feature for both language and music, but the way that rhythm is realized in each domain—that is, how rhythm unfolds in time—is different.

Rhythm, in both music and language, can be defined as the pattern of ‘events’ in time (McAuley, 2010; Ravignani and Madison, 2017). Events in language typically occur at the syllable level, and events in music occur at the note level. Music and language differ in how the time intervals between events are structured. In musical rhythms, events are usually structured around a beat, or an underlying pulse (Drake, 1998; McAuley, 2010). Even though individual events are not equally spaced, the intervals between events relate to the beat, which means that durations are most commonly related by small integer ratios like 1:2 (e.g., quarter note:half note). The beat in music leads to the perception that the intervals between beats are roughly the same duration (i.e., isochronous; Ravignani and Madison, 2017; Ravignani and Norton, 2017) and gives listeners the sense of periodicity, or the perception of a pattern repeating regularly at a fixed period or interval in time (Patel, 2003; Patel et al., 2005; Kotz et al., 2018). Periodicity is present in music even despite natural tempo fluctuations or expressive timing that make a strictly isochronous beat improbable in human produced music (Fraisse, 1982; Epstein, 1985; Bharucha and Pryhor, 1986). In contrast, speech rhythms do not have a beat. It is this presence of a beat that we call rhythmic regularity.

Despite a long history of searching for strictly periodic intervals at the syllable or stress level in speech, no one has found regularly repeating patterns of equal duration in speech (Grabe and Low, 2002; Patel, 2003; Patel et al., 2005; Cummins, 2012; Goswami and Leong, 2013; Brown et al., 2017). Although speech sounds are generally considered rhythmic, those rhythms are constrained to the length of the word, linguistic stress pattern, syntactic rules, or prosodic emphasis in a sentence (Cutler and Foss, 1977; Hay and Diehl, 2007; Turk and Shattuck-Hufnagel, 2013), which does not lend well to rhythmic regularity. These temporal regularities are crucial for speech intelligibility (Shannon et al., 1995) and more crucial than spectral characteristics of speech (Albouy et al., 2020). Speakers learn the typical rhythmic patterns of their language and this knowledge gives rise to temporal predictability in speech (Rosen, 1992; Hawkins, 2014; Jadoul et al., 2016; Rathcke et al., 2021), rather than any rhythmic regularities in the speech signal (Beier and Ferreira, 2018). The differences in regularity between music and language are especially salient when comparing sensorimotor synchronization to speech and song, where speech has much greater variability in the alignment of taps to syllable events in speech (30%) compared to note events song (4%, Lidji et al., 2011; Cummins, 2012; Dalla Bella et al., 2013).

In each domain, there is considerable research characterizing the degree or type of rhythmic information in the signal. These studies ask, for instance, whether language is rhythmic at all (e.g., Nolan and Jeon, 2014) or what acoustic factors contribute to the strength of perceived regularity in music (e.g., Bouwer et al., 2018). A range metrics have been used to characterize rhythm and/or regularity within each domain and, in a few cases, across domains. These metrics include the calculation of inter-onset-intervals between successive notes or syllables (e.g., stressed and unstressed IOIs; Vanden Bosch der Nederlanden et al., 2022a,b), durational contrastiveness between pairs of successive notes or syllables (Pairwise Variability Index; Grabe and Low, 2002; Patel and Daniele, 2003; Hannon, 2009; Hannon et al., 2016), the proportion of vocalic intervals in an utterance (vowel reduction; Grabe and Low, 2002; Wiget et al., 2010; Arvaniti, 2012), acoustic feature extraction using music information retrieval techniques (e.g., Lartillot and Toiviainen, 2007; Lartillot et al., 2008; Alluri and Toiviainen, 2010; Burger et al., 2013, 2014), autocorrelations to detect self-similarity in the envelope of a signal (Leong, 2012; Suppanen et al., 2019), clock timing evidence and counter-evidence (Povel and Essens, 1985), and integer multiple relatedness (Roeske et al., 2020; De Gregorio et al., 2021). These metrics have been useful within their own contexts of identifying, for example, whether a composer’s language background influenced the musical rhythms they employed (Patel and Daniele, 2003; Van Handel, 2006) or determining the strength of a beat in one musical rhythm compared to another (Henry et al., 2017; Matthews et al., 2019). However, not all speech-rhythm metrics have proven to be reliable or strong predictors of perceived speech rhythms (White and Mattys, 2007; Arvaniti, 2012; Jadoul et al., 2016). In music, the task of beat extraction is difficult (McKinney et al., 2007; Grosche et al., 2010), even if humans do it spontaneously (Grahn and Brett, 2007; Honing, 2012). The goal of the current paper is to examine whether some of the above metrics used to characterize rhythmic regularity in music or language separately can characterize the differences in rhythmic regularity between language and music.

Past work has examined where in the acoustic signal the beat is located in speech and song, finding consistent tapping in speech and song at p-centers (but see conflicting takes on p-centers Morton et al., 1976; cf. Marcus, 1981; Vos and Rasch, 1981; Pompino-Marschall, 1989; Scott, 1998; Villing et al., 2007), vowel onsets (Rathcke et al., 2021), or at peaks in the acoustic envelope (Kochanski and Orphanidou, 2008; Lartillot and Grandjean, 2019). Still others have used cochlear models of acoustic salience to find the beat location in vocally-produced songs (Ellis, 2007; Coath et al., 2009). While these approaches are germane to the current question, our goal is to determine whether acoustic features of speech and song can eventually provide evidence of rhythmic regularity—in the form of an equally-spaced, repeating pulse—in a range of communicative and non-communicative domains. For instance, there is increasing evidence that regularity is a salient feature in the sensory landscape (Aman et al., 2021), with listeners detecting regularity within a single cycle of it emerging from a random background (Southwell and Chait, 2018) or preferentially attending to a visual stream with statistical regularities despite having no conscious perception of that regularity (Zhao et al., 2013). Stimuli in studies like these are created with careful control over what features should give rise to regularity, but a wide range of natural stimuli, including non-human animal vocalizations (Kotz et al., 2018; Roeske et al., 2020; De Gregorio et al., 2021) and environmental sounds (e.g., Gygi et al., 2004; Rothenberg, 2013) also give rise to regularity in a variety of different acoustic characteristics. Our goal is to find a metric that indexes the differences in regularity between speech and song with the future goal of using this metric to detect the degree of regularity in a range of naturally occurring sounds.

Acoustic features that differentiate temporal regularity in speech and song will also feed into perceptual and cognitive questions related to how humans differentiate speech and song in development (Vanden Bosch der Nederlanden et al., 2022a,b). Rhythmic regularity is an important feature for speech-to-song or environmental sound-to-song transformations (Simchy-Gross and Margulis, 2018; Tierney et al., 2018; Rowland et al., 2019), but spectral features seem to be better predictors of a listeners’ perception of an utterance as speech or song (Hilton et al., 2022; Vanden Bosch der Nederlanden et al., 2022a,b; Albouy et al., 2023; Ozaki et al., 2023). Given the importance of rhythmic differences between and among languages for helping children acquire language (Ramus et al., 1999; Nazzi et al., 2000; Jusczyk, 2002), and for bringing about a transformation from speech to song, a clear acoustic metric of rhythmic regularity may prove useful for understanding the development of distinct domains of communication.

We address the goals in the current study by first obtaining subjective ratings of the differences in rhythmic regularity between spoken and sung utterances. After establishing this subjective metric, acoustic features of spoken and sung utterances were related to subjective ratings of rhythmic regularity to examine which features are most predictive of perceived rhythmic regularity.

Experiment 1 illustrated that participants had varying definitions of rhythmic regularity when we left it undefined and did not provide training examples. Initially it appeared that our acoustically matched stimuli did not differ in perceived rhythmic regularity, but after taking participants’ definitions into account (whether their definition was beat-based, normal-prosody, or unclear), regularity was greater for song than speech for beat-based definitions and greater for speech than song for normal-prosody definitions. Note that the normal-prosody definition group did not describe prosodic rhythmic regularity or a beat in speech, but rather participants in this group largely based their definitions only on the regular part of the term rhythmic regularity. Instead, these participants focused on how normal the speech sounded for everyday conversations. Although definition groupings explained a significant amount of variability in regularity ratings, it is also possible that the acoustic constraints placed on the stimuli reduced the differences in rhythmic regularity between spoken and sung exemplars. In this case, different profiles of regularity for speech and song in Experiment 1 may mean stimuli did not differ or only weakly differed in rhythmic regularity. We designed Experiment 2 to determine whether providing a clear definition of rhythmic regularity would shift participants’ ratings to align with the beat-based definition of rhythmic regularity we set out to examine in addition to determining whether regularity ratings were consistent across different stimulus sets.

We improved on Experiment 1 in three ways: (1) We provided a concrete rhythmic regularity rating scale “How easy would it be to tap or clap along to that clip?” (2) We provided training examples before participants began the rating task consisting of spoken and sung clips that would be easy and not easy to tap or clap to using familiar stimuli, and (3) We added a second unmatched stimulus set of spoken and sung stimuli that were not acoustically matched to examine regularity differences between unconstrained spoken and sung exemplars.

A second goal of Experiment 2 was to relate participants’ regularity ratings to acoustic features of spoken and sung exemplars. To achieve this goal, speech- and music-based acoustic features were extracted from all stimuli using Praat, MIR Toolbox, and custom music-inspired scripts (see OSF). We used standard acoustic features that are known to differ between speech and song (Vanden Bosch der Nederlanden et al., 2022a,b), as well as several features described in the introduction related to temporal regularity (see Appendix D for full feature list).

A major goal of Experiment 2 was to standardize participants’ interpretation of rhythmic regularity by providing a concrete definition centered on ease of clapping or tapping along with the stimulus. With this definition, rhythmic regularity ratings were significantly higher for sung than spoken utterances. Experiment 2 also expanded on the acoustically matched stimulus set from Experiment 1 by including an additional unmatched stimulus set more representative of speech and song in everyday settings. Participants rated song as more rhythmically regular than speech for both sets, but the difference was larger for the acoustically unmatched than the matched set. Naturally recorded utterances may emphasize the differences in regularity between song and speech compared to recordings that equate tempo, pitch contour, and average pitch between speech and song. However, regularity differences are apparent even in carefully acoustically matched stimulus sets, suggesting that regularity helps differentiate speech and song. Finally, we estimated which acoustic features across both stimulus sets were most predictive of regularity ratings. Although the type of stimulus (speech or song) was a significant predictor of regularity, longer syllable durations and less spectral flux also predicted higher rhythmic regularity ratings.

The goal of this work was to obtain a subjective metric of rhythmic regularity—an equally-spaced, repeating pulse—and examine acoustic features that predict participants’ ratings of regularity. Experiment 1 illustrated that the term rhythmic regularity was interpreted differently across participants, leading to different patterns of regularity across spoken and sung exemplars. Experiment 2 operationalized the definition of rhythmic regularity by asking how easy it would be to tap or clap to the stimulus. With this definition, participants rated song as more regular–or easier to clap or tap to–than speech in both acoustically matched and acoustically unmatched stimulus sets. Subjective regularity ratings were significantly affected by acoustic features of syllable duration and spectral flux, with longer durations and less flux related to higher regularity ratings. These results add to the literature by (1) highlighting the salience of rhythmic regularity as a differentiator of speech and song (Patel and Daniele, 2003; Patel et al., 2005; Vanden Bosch der Nederlanden et al., 2022b) and (2) adding to the growing literature on spectral flux as a salient acoustic feature in listeners’ perceptual processing of sound (Weineck et al., 2022).

Spectral flux is a metric of the distance between successive frames, or moments in time, in the frequency spectrum, with larger values indicating large changes in the spectrum from moment to moment (Alluri and Toiviainen, 2010). It logically follows that song should have less spectral flux since notes are held longer (i.e., greater proportion of the utterance is vocalic) than in speech, creating fewer changes in the spectrum on a moment-to-moment basis. The metrical framework of sung utterances may also make for fewer sudden and more evenly spaced changes in the spectrum compared to speech. Spectral flux has been described as an acoustic correlate of the beat in music, but with greater spectral flux indicating greater beat salience (Burger et al., 2013). These authors extracted spectral flux from low and high frequency bands in the spectrum corresponding to the kick drum, hi-hat, and cymbal. For this reason, large amounts of spectral flux in these bands acted as a proxy for rhythmic information from these instruments. These stimulus-specific differences help to explain the seeming paradox of greater spectral flux predicting more beat salience in music, while greater spectral flux predicts less rhythmic regularity when comparing speech to song.

Our results elucidate what features participants use to provide regularity ratings when comparing speech and song, but these features alone are unlikely to capture the presence of a beat or the integer multiple relatedness of sounds snapping to the metrical grid across a wide range of environmental stimuli. We attempted to account for listeners’ subjective regularity ratings using several music- and language-inspired metrics of regularity. In particular, the proportion of intervals per sentence that were related by integer multiples (Roeske et al., 2020) was not correlated with regularity ratings. It may be that our sentence-level approach is too coarse a metric and behavioral responses like tapping or continuous regularity ratings could shed light on which features participants relied on at particular moments in time to feel a beat (similar to Rathcke et al., 2021). The consistency with which those moments align with inter-onset-interval or stimulus features could provide a path forward for creating novel metrics to characterize regularity differences in speech and song. Another set of metrics used for this study (Asynchrony, Signed Asynchrony and their variability) was inspired by the clock timing work from Povel and Essens (1985) (similar to Norton and Scharff, 2016 for birdsong). However, this metric also failed to provide any relationship to subjective regularity and may also require input from the p-center-related literature (e.g., Rathcke et al., 2021) to determine the correct beat locations and onset times used to develop the underlying “clock” for speech and song. Onset intervals related to vocalic or other salient features of the stimulus may be more fruitful than the reliance on linguistic onsets used here. Finally, music information retrieval metrics like pulse clarity and stimulus-extracted tempo had no relationship to rhythmic regularity in speech and song, suggesting that these feature extraction methods are perhaps better suited for use with multi-instrument (e.g., vocals and instrumentation) excerpts of musical pieces rather than vocal sung and spoken utterances.

Linguistic measures, including measures that have previously been used to relate speech and music to one another, such as nPVI, also did not explain additional variance in rhythmic regularity beyond average syllable duration (see Supplementary Table S6). Vocalic nPVI was originally developed to capture the vowel reduction (i.e., change in vowel quality to a “schwa” and shortened duration of vowel length) that happens in many of the so-called “stress-timed” languages (Grabe and Low, 2002; Patel et al., 2005; Cummins, 2012). This measure is not best at capturing rhythmic variability, but rather contrastiveness between pairs of syllables. Indeed, our calculations indicated that music often had more contrastiveness than speech (see Table 1, Unmatched stimuli), which is likely due to large integer-related duration differences like quarter notes to half or whole notes that speech does not employ. Comparisons of previous work from separate studies suggested that nPVIs were much higher for speech (in the 50–70 range) than instrumental music (in the 30–40 range; Patel and Daniele, 2003; Hannon et al., 2016), but these studies used musical notation to estimate nPVI durations instead of actual recordings. Studies that have used acoustic segmentation of speech and song have illustrated more comparable nPVI values (Vanden Bosch der Nederlanden et al., 2022a,b). Thus, it is not surprising that this metric did not uniquely predict rhythmic regularity for spoken compared to sung stimuli.

Despite the ease with which humans pick up on regularity in speech, song, and environmental sounds, easily extractable acoustic features that characterize those subjective reports remain elusive. Our study confirms that participants hear more rhythmic regularity in sung compared to spoken utterances, providing concrete metrics for how best to obtain participant’s subjective regularity ratings. The findings from this study also add to the literature by characterizing that regularity is easier to detect–or more likely to be present–when syllables are longer, and when there is less moment-to-moment fluctuation in the spectrum. Future work should build on these results to develop more continuous and fine-grained metrics for quantifying rhythmic regularity from the acoustic signal. There is growing evidence that rhythmic regularity is an important signal for attention, perception, development, and movement (Grahn and Brett, 2007; Gordon et al., 2014; Bedoin et al., 2016; Trainor et al., 2018; Aman et al., 2021; Lense et al., 2021) in humans, and is present in a range of human and non-human primate communicative vocalizations (Roeske et al., 2020; De Gregorio et al., 2021), as well as many environmental sounds (Gygi et al., 2004). Indeed, the perception of rhythmic regularity is key to how both human and non-human animals (e.g., cockatoos, sea lions) align their movements to a beat (Fitch, 2013). A greater understanding of what acoustic features humans rely on to perceive regularity and extract an underlying pulse in communicative signals like speech and song will contribute to theories of evolutionary origins of beat processing (e.g., are the features humans use to find a beat the same or different from animals?) and theories about perceptual biases toward regularity in everyday soundscapes.

One potential limitation of the current study is the use of lyrics in both the music and language domains. We wanted to use speech and song because they exemplify the acoustic and structural differences between domains (Vanden Bosch der Nederlanden et al., 2022a,b), while maintaining the ability to control for timbral, semantic, and other temporal or spectral acoustic features. It will be important to characterize the role linguistic content plays in the perception of rhythmic regularity in song. For instance, is song without words perceived as more strictly regular that song with words given that note durations are less dictated by word length or stress? If so, then are instrumental melodies perceived as more rhythmically regular than songs without words? Or does linguistic or semantic content help to bolster temporal prediction for what type of note and/or word will come next? Similarly, would speech without semantic content (e.g., low-pass filtered) be perceived as more or less regular than semantic speech? This and future work will help shed light on the temporal features that distinguish speech and song and, more broadly, the domains of music and language.

The current findings add to the literature on rhythm in music and language by providing a concrete subjective metric of rhythmic regularity that reliably differs between speech and song across stimulus sets. The metric is simple to understand and can be used to characterize the perception of rhythmic regularity across developmental populations, in individuals with little or no musical training, and in a range of stimulus sets beyond music and language (e.g., bird song). Our findings are important for characterizing the inherent differences in music and language that (1) may be important for learning to differentiate musical and linguistic communication early in development (Vanden Bosch der Nederlanden et al., 2022a,b) and (2) underlie many of the perceptual advantages ascribed to music over language. For instance, cross-culturally humans prefer simple integer ratios in music (Jacoby and McDermott, 2017) and remember these musical rhythms better than syncopated rhythms that disrupt the occurrence of events on a beat (Fitch and Rosenfeld, 2007). Future work comparing the prominence of features in speech compared to song could address the divergence of musical and linguistic communication in humans. For instance, does the preservation of rhythmic regularity in music come at a cost to the transmission of quick messages meant to transact information? Is strict isochrony better for promoting verbatim memory of information occurring on, but not off the beat (Jones et al., 1981; Large, 2008; Helfrich et al., 2018) while vague periodicity without strict isochrony (as in speech) is better for encoding the gist of a message? Answering seemingly simple questions like how humans perceive differences in rhythmic regularity in speech and song, has the potential to address several important areas of psychology related to human communicative development, origins of music and language, cross-species comparisons, and perceptual biases toward regularity in everyday scenes.

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found at: https://osf.io/hnw5t/.

The studies involving human participants were reviewed and approved by the University of Western Ontario Ethics Board. The patients/participants provided their written informed consent to participate in this study.

CY, JG, and CV designed the experiments. CY and CV recruited the participants and performed the data analysis. CV and AC extracted acoustic features and manually segmented stimuli. CY wrote the first draft. CV provided subsequent drafts. All authors contributed to the article and approved the submitted version.

This work was supported by NSERC RGPIN-2016-05834 awarded to JG and NSERC RGPIN-2022-04413 and DGECR-2022-00294 awarded to CV.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.