As learning and memory functions decline with normal ageing, the present study used age-balanced naturalistic stimuli to prevent bias from memorisation of study materials. Candidate items for the Repetition and Completion tasks (40 proverbs +40 songs) were pre-selected for a pilot questionnaire from collections of commonly known proverbs and songs. Candidate selection prioritized items of Finnish origin to avoid interference from other language versions and was restricted in length to minimise variability in the final stimulus set. Syllable count was used as the main measure of length because compound words and suffixes are typical for Finnish language: using word count might lead to remarkable differences in phrase length. The song material comprised old children’s songs and nursery rhymes, folk songs, evergreens, Finnish pop/rock classics, and seasonal songs. Song phrases were selected either from the beginning of the first verse or chorus, avoiding internally repetitive, non-unique phrases and non-lexical utterances such as “ooh.”

The familiarity of all candidate items was tested with an online pilot questionnaire using a 4-point Likert scale (1 = completely unfamiliar, 4 = extremely familiar). Specifically, respondents were instructed to report their personal feeling of familiarity with each item, emphasising that the questionnaire was not a test but aimed at providing information about which proverbs and song phrases were most and least familiar at the population level. There were also five fabricated control items (non-existing proverbs/song phrases) in each category to validate the reliability of the ratings. To control for potential age differences in familiarity, the respondents of the pilot questionnaire (N = 665) were divided into three age groups (20–39 years, N = 246; 40–59 years, N = 294; 60 and above, N = 125) and 20 proverbs +20 song phrases per task were selected from the items rated most familiar across all age groups.

Items constructed for the Improvisation tasks were carefully matched to the selected Completion task items in phrase length, sentence structure, semantic theme, as well as for musical style in the Song Improvisation task. Selected items were recorded in female and male voice, matching all items in duration, tone, and expression. Recordings were carried out by a professional female and male musician who also composed the melodies for Song Improvisation items.

Stimuli consisted of familiar proverbs and song phrases for Repetition, beginnings of familiar proverbs and song phrases for Completion, and beginnings of new proverbs and song phrases for Improvisation; a total of 20 proverb +20 song phrase items in each task. After familiarity-based final selection, phrase lengths across both modalities remained similar, with song phrase items compared to proverb items being longer by 0.5 words / 1 syllable across all tasks (see Table 1). Melodic production in singing resulted in longer durations compared to speech. In addition, Completion and Improvisation items, containing only the first half of a phrase, were shorter than Repetition items which contained the full phrase.

Recruited participants reported their familiarity with memory task stimuli (Repetition items and cue parts of the Completion items) as a part of a comprehensive questionnaire battery assessing demographic information, health, and wellbeing as well as personal history with various leisure activities, including music (listening, singing, playing, and dancing). Although memory effects as such were outside the scope of this study, familiarity ratings were nevertheless expected to cause some memory activation. Participants were thus allowed to review the questionnaire items before scanning to balance out inter-individual differences from variable rating-scanning delays.

Prior to scanning, participants received instructions on a computer screen along with three practice trials mimicking the course of a task in the fMRI experiment. The researcher ensured that the participant understood the tasks and was able to carry them out. During scanning, the researcher informed the participant about which of the six tasks was about to begin and ensured that the instructions for the task were clear. For each task, participants were instructed to try their best and to always speak in the Proverb tasks and sing in the Song tasks.

Task order was pseudo-randomised in such a way that neither modality was presented twice in a row, no task type (Repetition / Completion / Improvisation) was presented twice in a row, and the first task rotated. Task order was counter-balanced between young (aged 20–39 years), middle-aged (40–59) and old (60–90) participants and between participants with regular (at least once per week for at least 1 year) vs. no singing activities. Stimuli were presented in random order within each task.

Data were acquired at the Advanced Magnetic Imaging (AMI) Centre of Aalto University, Espoo, Finland, with a MAGNETOM Skyra 3.0 T scanner (Siemens) using a 32-channel RF receiving head coil. Participant’s head was supported in place with soft foam paddings inside the coil to reduce head movement and protect hearing. Participants were asked to report immediately should they experience any discomfort or difficulty in perceiving instructions or stimuli. A whole-brain T1-weighted anatomical volume for anatomic normalisation was obtained using a 3D magnetisation-prepared rapid gradient-echo (MPRAGE) sequence: echo time (TE) = 3.3 ms, repetition time (TR) = 2530.0 ms, flip angle (FA) = 7 degrees, voxel size = 1 × 1 × 1 mm³, FOV = 256 mm and slices per slab = 176. T2*-weighted functional imaging was carried out using gradient echo-planar imaging (EPI) on the whole brain: TE = 30.0 ms, FA = 90 degrees, repetition time (TR) = 6,800–10,800 ms, delay in TR = 6,110–10,110 ms, voxel size = 3 × 3 × 3 mm³, FOV = 240 mm, and slices per volume = 44.

Before each measurement, sound volume was adjusted to a clearly audible but comfortable level, and microphone position was adjusted as per sound quality tests as needed. Visual instructions for active conditions were presented as graphical icons on a back projection mirror screen during tasks (see Figure 1). Auditory stimuli were presented using MR-compatible high-quality binaural insert earphones (KAR ADU2a). Vocal responses were recorded with a high-dynamic-range noise-cancelling optic microphone (Optoacoustics FOMRI-III) attached to the head coil. The tasks were created and instructions and stimuli were presented with Presentation 20.3 (Neurobehavioural systems, Inc., Berkeley, CA, United States) using MR pulses as triggers.

Each task involved 20 visually cued active condition triads: (i) listen and exhale, (ii) respond, and (iii) exhale (baseline), with inhale periods in between. A sparse-sampling technique (Hall et al., 1999; see also Belin et al., 1999; Yang et al., 2000) was adopted with silent intervals between consecutive whole-brain volume scans (see Figure 1). Each trial began with a long silent period (900–1,200 ms), followed by the active condition (length dependent on task), a short silent period (400–700 ms) and volume acquisition time window (1,500 ms). Since stimulus durations varied as per experiment design, times of repetition (TRs) were optimised task by task, resulting in TR = 6,800 ms in Proverb Repetition, 7,000 ms in Proverb Completion, 7,000 ms in Proverb Improvisation, 10,800 in Song Repetition, 8,800 ms in Song Completion and 8,000 ms in Song Improvisation. Systematic variation (jittering) of the timings of the active trials aimed for acquiring the peak of the hemodynamic response, estimated to occur with a 4–7 delay (Belin et al., 1999; Yang et al., 2000). Each task comprised 20 blocks of trial triads plus a start trial involving the initial inhalation, 61 trials in total.

Functional MR volumes were pre-processed with Statistical Parametric Mapping (SPM) 12 software package (The Wellcome Centre for Human Neuroimaging, UCL, London) running under Matlab 2019a (The MathWorks Inc., Natick, MA, United States). EPIs were realigned to the first scan by estimating the parameters of an optimal rigid body transformation. The mean and individual EPIs were also manually reoriented to the anterior commissure to improve co-registration with the T1-weighted (T1w) image. After co-registration of the T1w with the mean EPI as reference, segmentation was performed using unified segmentation (Ashburner and Friston, 2005) involving medium regularisation and light clean-up with SPM12 IXI data set tissue probability maps. All EPIs were normalized to the MNI template by applying the spatial normalisation parameters from the T1w image and resampling to 2-mm isotropic voxels using trilinear interpolation. As vocal production tasks were expected to cause head movement above that of an average tb-fMRI experiment, the Artifact Detection Toolbox (Mozes and Whitfield-Gabrieli, 2011; available at http://www.nitrc.org/projects/artifact_detect/) was run on the unsmoothed EPIs to identify potential outlier scans using a 3-mm framewise displacement threshold with global BOLD signal change threshold of 4 SD. Participants displaying a within-task outlier scan rate of 20% or above were rejected. Finally, EPIs were spatially smoothed using spatial convolution with an 8-mm FWHM Gaussian kernel to minimize effects of inter-individual anatomical differences.

Main effects were estimated with the general linear model (GLM) framework. Each condition was modelled with a boxcar function convolved with the canonical hemodynamic response function (HRF), as this has been shown to yield the best model fitting in sparse-sampling acquisitions (Perrachione and Ghosh, 2013). Based on previous literature (Belin et al., 1999; Kleber et al., 2010), we opted for a ‘slow’ event-response analysis and set trial duration to 0 s. Microtime resolution was adjusted to 125 ms per bin corresponding to the standard value used in continuous fMRI, with middle slice as the reference. A high-pass filter with a 128 s cut-off was applied to remove slow signal drifts. Temporal autocorrelations in the BOLD signal time series were accounted for with an AR (1) model to approximate the observed covariance of the functional data in the context of restricted maximum likelihood estimation. We included the TPM for grey matter tissue thresholded at 15% probability as explicit mask. No global scaling was applied. In addition to experimental condition effects, the design matrix included nuisance regressors to control for the effects of head motion (6 realignment parameters) and potential outlier scans (one regressor per outlier scan; Jenkinson et al., 2002). The model was first estimated within each subject, and contrast-of-interest volumes were generated comparing the appropriate conditional coefficients to assess the main effects of production versus exhalation, which served as the baseline condition.

Performance in the fMRI tasks was assessed in terms of response timing (all tasks) as well as stimulus familiarity ratings and response match with target (Repetition and Completion tasks) based on measurement documentation (filled-in log sheets) and recorded vocal responses. Given that the sample consisted of nonmusicians, with approximately half reporting no singing experience, and considering that singing skill and performance were not the focus of this study, we placed a greater emphasis on verbal accuracy over melodic accuracy when assessing sung responses in the Repetition and Completion tasks. In these tasks, deviations from target responses were marked as erroneous, while allowing for non-lexical utterances such as ‘um’ or ‘haha’ as fillers and replacing words with synonyms. To harmonize the rising of the BOLD responses across trials, timing was marked as erroneous should the response begin too early or end too late with more than 25% of the response falling outside the production time window. A technical failure in the laboratory resulted in poor audio recording quality for 26 participants, due to which errors documented during measurements could not be verified from recorded vocal responses. For these participants, performance was only assessed with measurement documentation.

Group effects were studied with (i) main effects across the age range, (ii) ageing effects, and (iii) ageing effects on laterality index (LI). First, as the paradigm allowed controlling for inhalation-related activation implicitly, production versus baseline (exhale) contrasts (Repetition > Baseline, Completion > Baseline, Improvisation > Baseline) from each participant were used directly to calculate the main effects of each task with one-sample t-tests. Direct modality comparisons (Proverb > Song, Song > Proverb) within each task type (Repetition, Completion, Improvisation) were calculated with paired t-tests. Ageing effects for speech, singing, and their within-subject difference were studied with multiple regression in both directions (i.e., increasing or decreasing with advancing age) for the whole brain. All models controlled for variation in education years, age-relative experience with musical hobbies (% of age active in musical hobbies in years) and total intracranial volume (TIV). These were supplemented with age for studying the main effects of each task (t-tests).

In addition, age-related laterality shifts [as predicted by the Hemispheric Asymmetry Reduction in Older Adults (HAROLD) model; (Cabeza, 2002)] were studied independently with 25-percent trimmed mean scores from bootstrapped laterality indices, a histogram-based fMRI adaptation of the classical approach

(LI toolbox for SPM; Wilke and Schmithorst, 2006; Wilke and Lidzba, 2007), on the whole brain (grey matter) with stepwise multiple linear regression in SPSS (IBM SPSS Statistics 28). Unfortunately, as the calculation of LI aborted if the minimum bootstrapped (within-subject) sample size was not reached, group-level sample sizes were partially reduced in LI analyses. In addition to whole-brain analyses, laterality changes in pre-defined regions of interest (ROIs), derived from the probabilistic connectivity-based Brainnetome atlas (available at https://atlas.brainnetome.org), were analysed for task pairs where age-associated functional reallocation showed a significant laterality change at larger scale. To test for the previously established laterality shift in the (pre-) frontal regions (Cabeza, 2002; see also Wierenga et al., 2008; Hoyau et al., 2017) and whether such an effect would also appear in temporal and parietal auditory-motor vocal production regions (Callan et al., 2006; Özdemir et al., 2006), this fine-grained approach involved MFG (Brainnetome subregions 15–28), IFG (29–40), orbitofrontal (OFG; 41–52) and precentral gyri (PrCG; 53–64) frontally; STG (69–80), middle temporal gyri (MTG; 81–88), inferior temporal gyri (ITG; 89–102), posterior superior temporal sulci (pSTS; 121–124) temporally; and inferior parietal lobules (IPL; 135–146) and postcentral gyri (PoCG; 155–162) parietally. In the stepwise regression, education years, musical hobbies, and TIV were introduced as independent variables in addition to age. Alpha levels (0.05) were corrected for multiple comparisons at each level of testing (Bonferroni).

Compared to singing, activity in left/bilateral PoCG and PrCG decreased with advancing age in Repetition and Completion of familiar proverbs. Consistent across all Proverb tasks, this difference was likely driven by speech. Centred in the somatosensory regions, such activation reductions might reflect reduced ability to process somatosensory feedback to support vocal motor production. It should also be noted that no lateralisation changes were found in the somatomotor regions, which suggests that these decreases are not characterised by hemispheric asymmetry reduction.

In favour of this account, PoCG and adjacent regions belong to typical sites of accelerated grey matter (GM) loss, often displaying left bias (Di et al., 2014; Minkova et al., 2017; see also Solé-Padullés et al., 2009).Consistent with the present results, a consequence from this interpretation is that speech, relying more heavily on left hemisphere regions potentially showing faster age-related atrophy, would be subject to more abrupt adaptation than singing. This is not to say that ageing impacted the modalities selectively but, rather, that the differential distribution of singing networks would provide neural resources for more flexible (internal) adaptability. Indeed, a recent study on post-stroke aphasia (Pitkäniemi et al., 2023) reported that singing shares language network resources with speech but relies less on the dorsal pathway typically associated with sensory-motor mapping (e.g., Saur et al., 2008). Flowing across the somatomotor regions, this distinction might explain a more rapid age-induced need for reorganisation in speech than singing. In their controlled study on primary progressive aphasia (PPA), Grube et al. (2016) also speculated that auditory deficits across different subtypes of PPA may relate to processing along the dorsal stream, which might in turn help explain previous observations of dissociated speech and singing ability in PPA (e.g., Polk and Kertesz, 1993; see also Baird and Thompson, 2019). Future studies should aim to map the network characteristics during singing which may overcome processing deficits impairing speech in clinical populations and, potentially, delay the need for network reorganisation in healthy ageing. Moreover, understanding both pathological and age-associated network changes in speech versus singing could directly benefit clinicians designing rehabilitative strategies for aphasias such as PPA.

As the singing network is inevitably subject to structural declines in healthy ageing, another mechanism to consider is that neural differentiation, i.e., dispersion from specialised to more general-purpose networks following GM atrophy (see Shafto and Tyler, 2014), might take a different trajectory in singing. Thus, a secondary suggestion for future work is to assess whether the somatomotor network’s use of external connections (e.g., Geerligs et al., 2015; see also Carp et al., 2011), for instance, varies by higher vocal motor control demands involved in the melodic-lexical co-processing in singing.

The sole ageing effect in singing was observed in Completion of song phrases from memory as increased activation in the medial frontal and parietal default mode network (DMN) regions, as well as bilateral MFG/SFG, left ITG/MTG, and right cerebellum. While these regions have been associated with recognition of familiar musical excerpts (Platel et al., 2003) and internal tasks such as retrieval of episodic and semantic knowledge (Smallwood et al., 2021), their increased activation in Completion but not in Repetition of familiar song phrases, specifically at an older age, seems to suggest a task-related effect.

Increased coupling of lateral prefrontal and DMN regions is thought to support cognitive performance by means of shifting from fluid abilities to higher reliance on experience (Turner and Spreng, 2015; Spreng and Turner, 2019), while increases in lateral cerebellar activation may support motor preparation for producing verbal content (Riecker et al., 2000; Tremblay et al., 2017). Indeed, this task elicited the highest number of average errors regarding a mismatch between target and response (Table 2). As the pattern did not appear in the singing versus speech comparison, this is suggestive of increased cognitive effort in retrieving a song phrase ending rather than representing an ageing effect related to singing or retrieval as such. Indeed, phrases belonging to the larger context of a song might be less automated than proverbs, which are complete linguistic entities on their own.

Ageing elicited increasing activation in right FG and posterior ITG and bilaterally in the AG with right-hemisphere bias in improvisation of new proverbs. These regions are typically associated with semantic tasks in the young adult brain (Benedek et al., 2014; Bonilha et al., 2017; Liu et al., 2021). Consistent with the task demands of Proverb Improvisation, age-related volume losses in FG and ITG have been associated with declines in verbal fluency, verbal memory, attention, and executive function (Armstrong et al., 2020; see also Spreng and Turner, 2019). Recent work (Adnan et al., 2019a,b; Martin et al., 2023) also suggests some involvement of FG and posterior ITG as a part of large-scale network reorganisation supporting semantic/creative cognition at older age, where bilateral FG may expand their communication network while other regions serve as additional connector hubs (Martin et al., 2023). Speculatively, the observed activation increases might thus represent successful functional compensation in response to cognitive task demands and ongoing network reorganisation.

In the Improvisation tasks, difference between the vocal modalities (speech > singing) showed an age-associated increase in activation in the right IFG (opercularis, triangularis) with a respective reduction of left bias in the laterality (LI) of IFG activity. These effects were potentially driven by a marginally significant increase of right IFG activation in speech—a classic example of prefrontal hemispheric asymmetry reduction in older adults (Cabeza, 2002) previously observed during rapid naming, for instance (Hoyau et al., 2017). Separate analyses of laterality confirmed minor rightward shifts across the whole brain and in the prefrontal regions such as MFG with older participants showing relatively increased use of right-hemisphere homologues in speech. By contrast, age was not associated with laterality in singing. Providing support for our last hypothesis, these results suggest that, under similar task demands, speech and singing show dissociative ageing effects on prefrontal hemispheric asymmetry.