Song Familiarity Relies on Evidence Accumulation

ABSTRACT

Familiarity judgments are thought to involve evidence accumulation, a decision‐making process in which information is gathered over time until a threshold is reached. Previous work has identified a scalp‐recorded signature of evidence accumulation called the central‐parietal positivity (CPP). We built on this previous work and recorded electroencephalography while participants listened to several melodies, instructing participants to respond as soon as the song felt familiar. A prominent CPP was noted, time‐locked to decisions. We then used linear regression to unmix overlapping neural responses and observed a stimulus‐locked indicator of evidence accumulation that increased in amplitude just prior to a familiarity decision. This result suggests that song familiarity relies on evidence accumulation, with individual notes in a familiar song acting as “evidence”.

Short abstract

Our work shows that EEG can be used to measure decision‐related activity in naturalistic tasks. We used a regression‐based approach to unmix overlapping stimulus‐locked and response‐locked activity while people listened to music. Familiar songs evoked separable signals consistent with evidence accumulation, a decision‐making process previously studied in highly controlled perceptual tasks. This suggests that familiarity does not occur suddenly but rather relies on the gradual accumulation of information over time.

Familiarity is the feeling that we've encountered a stimulus previously and is often contrasted with recall, the retrieval of specific information from a past event. The relationship between familiarity and recall has been debated, but both processes seem to be important in recognition memory. According to some accounts, recognition judgments rely on both the “qualitative” recollection of details and the “quantitative” feeling of familiarity (Yonelinas 2002). Familiarity often occurs in the absence of recall and “recognition without identification” has been observed for words (Cleary and Greene 2000; Peynircioǧlu 1990), images (Curran and Cleary 2003), and songs (Kostic and Cleary 2009).

One possible mechanism underlying familiarity judgments is evidence accumulation, a decision‐making process by which information is gathered over time until a threshold is reached. Here, familiarity is viewed as a graded signal which, if large enough to exceed a criterion, leads to a familiarity judgment (e.g., a “know” response in a remember/know task: Tulving 1985). In other words, the feeling of familiarity builds over time rather than happening instantaneously. Evidence accumulation models are found in many areas of cognitive psychology, including perception, learning, language, value‐based decision‐making, and memory (Evans and Wagenmakers 2019). What counts as “evidence” in an evidence accumulation model varies by task and may include observables such as the direction a dot moves in a random dot kinematogram, or unobservables such as information retrieved from memory. For example, according to Ratcliff (1978), memory retrieval works when a stimulus is compared against several items in memory, with each comparison initiating a separate accumulation process. These processes terminate when a criterion is reached, leading to either a “match” or “non‐match” outcome.

Previous work has found convincing behavioral evidence that recognition memory relies on evidence accumulation. In particular, evidence accumulation models do an excellent job of explaining patterns in accuracy and response time (Evans and Wagenmakers 2019; Osth et al. 2018; Ratcliff 1978). In contrast, neural evidence for evidence accumulation during recognition memory has been mixed. For example, van Vugt et al. (2016) recorded electroencephalography (EEG) while participants made old/new judgments in a face recognition task. ¹ To identify the neural correlates of evidence accumulation, van Vugt et al. (2016) used regression to locate ramping activity occurring between a memory probe and a participant's response. They focused their analysis on several brain oscillations, but did not find any ramping signals, which they interpreted as evidence against a role for evidence accumulation in recognition memory. However, van Vugt et al. (2019) later identified a ramping stimulus‐locked ERP component called the central‐parietal positivity (CPP), which they interpreted as indicating an evidence accumulation process during recognition memory. Additionally, they related the observed signal to “strength of evidence”, which they defined as the difference between the current face and the remembered face. Their results align with previous work showing that the CPP is a neural marker of perceptual evidence accumulation (Devine et al. 2019; Dou et al. 2024; Kelly and O'Connell 2013; Loughnane et al. 2018; O'Connell et al. 2012; Pisauro et al. 2017; Ruesseler et al. 2022; Twomey et al. 2015, 300).

Prior work by van Vugt et al. (2016, 2019) is conflicting and has two limitations. First, it is unclear whether the central‐parietal activity they observed was related to the appearance of the memory probe, the subsequent decision/motor response, or both. Second, they used a delayed match‐to‐sample task, which may have included both a recollection component and a familiarity component. Whether or not familiarity alone elicits a neural measure of evidence accumulation is currently unknown.

To address these limitations, we recorded EEG while participants completed a song familiarity task. We focused on song familiarity under the assumption that individual notes could be treated as observable evidence in favor of a familiarity decision. ² Participants listened to song melodies and indicated familiarity with a button press. For familiar songs, participants were asked to identify the song. Thus, although we did not have precise control over the strength of evidence each note provided, ³ we planned to test the effect of “strength of evidence” by contrasting stimulus‐locked activity for unfamiliar notes and familiar notes (i.e., those preceding a familiarity judgment).

Our logic in focussing on individual notes, as opposed to the whole song, was as follows. First, we assumed that an accumulation signal would increase over time as each successive note contributed evidence in favor of a familiarity decision. Second, we speculated that although this accumulation signal would not be time‐locked to any single note, its effects ought to be detectable nevertheless in the note‐locked response as a positive‐going deflection at central‐parietal sites (i.e., it ought to resemble the CPP to which it contributes)—especially when familiar and unfamiliar notes are contrasted. Supporting this prediction, previous work has found that sequential events contributing evidence to a decision can themselves elicit a CPP‐like signal. For example, Ruesseler et al. (2023) used a novel random dot motion task in which the strength of evidence (the coherence of the dots) changed continuously over time. They found that sudden changes in evidence elicited a positive‐going deflection at central‐parietal sites. Similar stimulus‐locked activity can be seen when participant view a sequence of Gabor patches before judging the average orientation (Wyart et al. 2012).

To address the issue of overlap between stimulus‐locked activity and response‐locked activity, we used a regression‐based approach called deconvolution (Ehinger and Dimigen 2019; Smith and Kutas 2015a, 2015b). This allowed us to examine stimulus‐locked activity and response‐locked activity in isolation. Based on previous work (O'Connell and Kelly 2021; van Vugt et al. 2019), and if familiarity processes rely on evidence accumulation, one or both of the following results should emerge. First, EEG time‐locked to familiar notes should be more positive at central‐parietal locations compared to unfamiliar notes (i.e., an enhanced stimulus‐locked CPP). Second, EEG time‐locked to familiarity decisions should be ramping and positive (i.e., an enhanced response‐locked CPP).

1. Method

1.1. Participants

No power analysis was conducted because we had no prediction about the amplitude of the CPP (familiarity‐related CPPs have not been measured previously). However, prior studies have observed a reliable CPP for other decision types for sample sizes ranging from 18 to 25 participants (Dou et al. 2024; Kelly and O'Connell 2013; O'Connell et al. 2012; Pisauro et al. 2017; van Vugt et al. 2019). We chose to test 30 participants with an average age of 22.63 years (SD = 5.38). Ten of the participants self‐reported as male; the rest self‐reported as female. Three participants self‐reported as left‐handed, one as ambidextrous, and the rest were right‐handed. Five of the participants had coarse curly hair and were included in the study but required a greater‐than‐average amount of electrolytic gel to lower electrode impedances. Participants provided informed consent and were compensated for their time with course credit. The study was approved by the Research Ethics Board at MacEwan University. One participant was excluded from the study due to technical issues with the EEG amplifier. Two additional participants were excluded due to low trial numbers (two or fewer familiar songs).

1.2. Apparatus and Procedure

Participants sat approximately 625 mm from a 527 × 296 mm display (60 Hz, 1920 by 1080 pixels, Elo 2402L). Visual and auditory stimuli were presented using Psychtoolbox (Brainard 1997; Pelli 1997) in MATLAB 2023a (Mathworks). Audio was produced using an audio interface (Focusrite Scarlett Solo 4th Generation) and two studio monitors approximately 750 mm from the participant's left and right ears (JBL 305P MkII). The audio was adjusted to a comfortable level, and participants were asked to minimize head and eye movements throughout the experiment.

Participants then completed a song familiarity task. At the start of each trial, a song melody would play. There were 121 melodies in total, which were taken from a previous study (Kostic and Cleary 2009) and consisted of piano notes only. The melodies were selected at random (without replacement) for each participant and ranged in duration from 4.08 to 14.91 s (M = 8.45 s, 95% CI [8.05, 8.85]). The number of notes in each song ranged from 12 to 64 (M = 28.34 notes, 95% CI [26.62, 30.23]). Participants were instructed to “Press the SPACEBAR as soon as the song feels familiar, regardless of whether or not you can name it.” While the song played, a small fixation cross was displayed on the screen; participants were instructed to keep their eyes at this location. If the participant pressed the spacebar to indicate that the song felt familiar, the display immediately switched to the text prompt “Name, lyric, source?” Participants were instructed that this was their cue to identify the name of the song, part of the lyrics, or the “song source” such as a show or movie where they had heard it previously. If participants could not identify anything about the song, they were instructed to leave the prompt blank. The prompt disappeared when the participant pressed the enter key and was replaced by four song titles, including the true song title. Participants were instructed to select the title of the song they had just heard by pressing one of the number keys “1” through “4”. If they selected the correct song title, the selection turned green, and a checkmark appeared; otherwise, the selection turned red, and an “x” appeared. For trials in which the song was unfamiliar to the participant, the song stopped playing after several seconds and the rest of the trial proceeded the same way (the same text prompt followed by the multiple‐choice query). See Figure 1 for a trial overview.

FIGURE 1.

Task details and behavioral data. Participants listened to several songs and made a familiarity judgment (button press for familiar songs, no response for unfamiliar songs). Participants were then given two opportunities to identify the song: A text prompt and a multiple‐choice question (with feedback).

1.3. Data Collection

For each song, the experimental program recorded the song title, whether the participant responded (and if so, the response time), the participant's response to the “Name, lyric, source?” prompt, and the participant's response to the multiple‐choice query. We also recorded 32 electrodes of EEG referenced to electrode Fz using Brain Vision Recorder (Version 1.26.0001, Brain Products GmbH). Thirty electrodes were placed on an elastic cap with standard 10–20 layout (EASYCAP GmbH). Two electrodes were attached to the left and right mastoids to serve as a later reference. Conductive gel was used to lower electrode impedances. The EEG was sampled at 1000 Hz and amplified using a BrainAmp DC (Brain Products GmbH) with a 250 Hz anti‐aliasing hardware filter.

1.4. Data Analysis

1.4.1. Behavioral Analysis

The behavioral analysis was done in MATLAB 2024b (MathWorks). The goal of the behavioral analysis was to check the feasibility of our event‐related EEG analysis. The CPP is similar to another component called the P300, which has been shown to require around 20 trials per condition (Boudewyn et al. 2018; Cohen and Polich 1997). For each participant and trial, we manually examined what was written in response to the “Name/lyric/source?” prompt. Trials with blank responses were labeled as unidentified, as were trials with statements such as “not sure” or “don't know”. All other responses were labeled as “identified”, regardless of correctness. For each participant, we then calculated the total number of unfamiliar songs and familiar songs. To inform our stimulus‐locked analysis, we calculated the total number of notes each participant heard for each song type (unfamiliar, familiar). We also calculated the proportion of each song type that could be identified using the text prompt and using the multiple‐choice prompt. As mentioned above, we did not check the text response for correctness. However, only correct responses to the multiple‐choice query counted as “identified”. Finally, for each participant, we also calculated the mean response time (RT) relative to the onset of each song and relative to the onset of the previous note event (see Figure S1 for the RT distributions).

1.4.2. EEG Preprocessing

The EEG was analyzed in MATLAB 2024a (MathWorks) using the EEGLAB library (Delorme and Makeig 2004). After downsampling to 250 Hz, we applied a bandpass filter (0.1–20 Hz, 60 Hz notch), and rereferenced to the average of the mastoid signals. Ocular artifacts were identified and removed using independent component analysis (ICA). The ICA was trained on the entire recording, minus any samples that were flagged as artifacts. These samples were identified using a sliding‐window approach with a 2000 ms window and a step size of 100 ms. If the voltage in the window dropped below −500 μV or exceeded 500 μV, all samples in the window were flagged and excluded from the ICA. Ocular components were then identified using the function iclabel and removed from the continuous data set if assigned an “Eye” label with a likelihood exceeding the sum of all “Non‐Eye” components (Pion‐Tonachini et al. 2019). On average, we removed 3.59 independent components (SD = 1.18).

1.4.3. Event‐Related Analysis

To unmix the stimulus‐locked and response‐locked signals, we used an approach called deconvolution, which models the ongoing EEG as a linear combination of underlying “regression ERPs” (Burns et al. 2013; Ehinger and Dimigen 2019; Smith and Kutas 2015a, 2015b). This approach yields a regression ERP (rERP) for each participant, electrode, and modeled event that could be subjected to the same statistics as an ERP, discussed below. Using a toolbox called Unfold (Ehinger and Dimigen 2019), we constructed a regression model that included a −1500 to 1500 ms window around unfamiliar notes, familiar notes, and motor responses. A large window was chosen to capture both stimulus‐locked and response‐locked activity (and Unfold does not allow analysis windows of different length). After constructing the regression (i.e., after generating the design matrix), we checked each participant's continuous EEG for artifacts using a 1000 ms sliding window (100 ms step size). Unlike a traditional ERP analysis, this procedure flagged individual samples, not entire epochs. If any sample within the window exceeded 75 μV or fell below −75 μV, all samples within the window were flagged as artifacts and excluded from the regression. This was done separately for each electrode; if more than 10% of modeled samples were flagged for a particular electrode then we re‐ran the preprocessing step for that participant after removing and interpolating the data from the bad electrode. The data was removed prior to the ICA step and interpolated afterward. On average we interpolated 0.27 electrodes per participant (SD = 0.59). One participant was excluded from further analysis due to excessive artifacts (17 of their electrodes violated our “10% rule”). In total, our EEG analysis consisted of 26 participants. We also constructed standard ERPs, the details and results of which can be seen in the Figures S2 and S3.

For each participant, we then constructed and solved a regression equation of the form:

$$\mathrm{EEG}=\mathit{X\beta }+\epsilon$$

where X represents the design matrix of regressors, β represents the beta coefficients (the rERPs), and ε represents an error term. The design matrix X was constructed using the Unfold functions uf_designmat and uf_timeexpandDesignmat. Together, these functions created a time‐expanded design matrix consisting of “stick functions” (ones on the diagonal) around each event. Thus, the design matrix X had rows corresponding to EEG samples (number of rows = number of recorded samples), and columns corresponding to the different timepoints of the estimated rERPs. See Figure S4a for a visualization of the design matrix. Rows corresponding to artifacts were removed prior to estimating the solution.

Regularization was used to prevent model overfitting. We used a first‐derivative form of Tikhonov regularization that penalizes the sample‐to‐sample change in the solution. This imposes a smoothness constraint on the solution (Kristensen et al. 2017; Reichel and Ye 2008). Ten‐fold cross‐validation was used to select the optimal regularization parameter for each participant. Error was defined as the average mean‐squared error across all electrodes. The following lambda values were tested: 0 (no regularization), 10, 100, 1000, 10,000, 100,000, 1,000,000, 10,000,000, 100,000,000. The optimal lambda for each participant minimized the mean fold error across all ten folds and was consistent across participants: λ = 10,000 (10 participants) or λ = 100,000 (16 participants). After regularization, the beta weights (the rERPs) were estimated using the MATLAB function pinv, which calculates the Moore‐Penrose pseudoinverse of a matrix.

1.4.4. Parametric Regression

The event‐related analysis described above has a potential flaw: It treats all note events equally, regardless of their position relative to the familiarity decision. It could be, for example, that the notes immediately preceding a button press are processed differently compared to the notes toward the beginning of the melody. This would align with previous work showing that musical notes are processed in the context of previously heard notes (Sankaran et al. 2024). To test this, we included note position relative to the button press as a parametric regressor in a second regression analysis. The model also included constant (intercept) terms for notes and button presses for familiar songs. Note position was defined as follows: the note immediately preceding the button press was assigned a position of −1, the note before that −2, and so on. Only notes and button press events from familiar songs were included in the model. See Figure S4b for an example design matrix. We then estimated the rERP for each regressor using the same procedure and regularization parameters as before.

1.5. Statistics

To test whether unfamiliar stimulus‐locked activity differed from familiar stimulus‐locked activity, we used cluster based permutation testing as described by Maris and Oostenveld (2007). For each electrode, we computed a repeated measures t‐statistic at each sample point to determine whether the familiar stimulus‐locked voltage at that sample point differed from the unfamiliar stimulus‐locked voltage. We then identified clusters of sample points for which the t‐values fell below the 2.5th percentile or exceeded the 97.5th percentile. Spatial clusters were defined according to a template from the FieldTrip toolbox (Oostenveld et al. 2010). For each spatial cluster, we defined a “cluster mass” as the sum of the absolute values of the temporal cluster t‐values; to be included in a cluster mass, the voltage at a sample point had to reach significance for all members of the spatial cluster. We restricted the stimulus‐locked analysis to a window from 0 to 600 ms relative to the onset of the note, a typical stimulus‐locked ERP time range. To determine whether the observed cluster masses exceeded what could occur by chance, we permuted the participant rERPs by randomly swapping the condition labels on a participant‐by‐participant basis. This is equivalent to randomly flipping the sign of a data point in a single‐sample t‐test. We then computed the cluster masses of the permuted waveforms and recorded the maximum cluster mass, or zero if no clusters were found. We tested 1000 permutations in total. Finally, we labeled an observed cluster as “significant” if its cluster mass exceeded 95% of the permuted cluster masses. The reported p‐value was the proportion of permuted cluster masses exceeding the observed cluster mass. For each significant cluster, we also reported an effect size by averaging the EEG over the cluster electrodes and sample points for each participant and computing:

$$\text{Cohe}{\mathrm{n}}^{\prime }\mathrm{s}d=\frac{M_c}{s_c}$$

where M _c and s _c represent the mean and standard deviation of the resulting cluster voltages.

A similar procedure was done for the response‐locked rERP. Our goal here was to determine whether (and when) the response‐locked rERP differed significantly from zero (an “existence test”). We did not split this analysis by note familiarity because responses were only made in the case of a familiar song. To capture decision‐related activity, the response‐locked analysis was restricted to a window from 1000 ms before the response 100 ms after the response. Permutation testing proceeded as before, except with a single‐sample t‐test (i.e., a repeated‐measures test against zero). For both the stimulus‐locked and response‐locked analysis, and in the case that a significant cluster was found, we visualized the spatial pattern of the signal by calculating the average rERP across the significant temporal clusters for each participant and electrode.

For our parametric regression, we used cluster‐based permutation testing to conduct two existence tests. First, we checked to see where/when the note‐position rERP differed from zero. If found, a significant cluster would show an effect of note position on stimulus‐locked processing. Second, we checked for non‐zero response‐locked activity, a replication of our previous response‐locked analysis. This was necessary because the inclusion of an additional regressor could have influenced the response‐locked result.

Finally, we replicated our parametric analysis but excluded identified familiar trials (trials for which the participant could identify the song at the test prompt). This was done to control for the possible effect of recall on stimulus‐locked and response‐locked activity. In other words, we wanted to be confident that our EEG results were driven by familiarity alone. This analysis required that we exclude two additional participants due to insufficient trial numbers (total N: 24).

2. Results

2.1. Behavioral Results

Summary statistics, shown in Table 1, show a reasonable number of “familiar” responses. The number of familiar responses ranged greatly across participants, however (Figure 2). We observed that 43% of familiar songs could be identified in writing with a 68% score on the multiple‐choice question. We also observed that in all conditions, performance on the multiple‐choice question was above chance level (25%). Taken together, our behavioral results suggest that participants completed the task as intended, though there were fewer familiar responses than predicted based on pilot testing.

TABLE 1.

Behavioral results showing the mean number of trials and notes, the mean proportion of songs that could be identified, the mean proportion of songs that were correctly chosen at the final multiple‐choice (MC) prompt, and the response times (RTs) relative to song start and previous note event. For familiar songs, only the notes prior to the button press are included in the number of notes.

	Unfamiliar		Familiar
	M	95% CI	M	95% CI
Number of songs	33.8	[27.8, 39.9]	38.1	[30.5, 45.6]
Number of notes	924.6	[768.0, 1081.1]	571.1	[500.7, 641.5]
Proportion identified—Text	0.02	[0.01, 0.03]	0.43	[0.33, 0.52]
Proportion identified—MC	0.42	[0.36, 0.48]	0.68	[0.62, 0.75]
RT from song start (s)	—	—	4.73	[4.40, 5.06]
RT from previous note (s)	—	—	0.63	[0.59, 0.67]

FIGURE 2.

Behavioral results indicate that the task was completed as intended. The number of familiar songs ranged greatly across participants (a). Nevertheless, there were many notes per condition (b). At the text prompt, participants were more likely to identify familiar songs compared to unfamiliar songs (c). Multiple choice (MC) performance was greatest for familiar songs but was above chance level (the dotted line) in both conditions (d). Error bars indicate 95% confidence intervals.

2.2. EEG Results

2.2.1. Regression ERPs

After using linear regression to model our EEG (Figure 3a) we found no significant clusters differentiating familiar stimulus‐locked activity and unfamiliar stimulus‐locked activity. We have illustrated the mean stimulus‐locked rERP for each condition (familiar, unfamiliar) at two scalp locations: frontal (FCz) and parietal (CPz)—see Figure 3b. This suggests that, on average, stimulus‐locked activity did not depend on familiarity (though see our second EEG analysis below). The response‐locked result showed activity in line with a CPP, and a significant cluster from −364 ms to 100 ms around O1 was identified (O1, Oz), p = 0.0017, Cohen's d = 0.71 (Figure 3c). See Figure S5a for individual participants' scores.

FIGURE 3.

Deconvolution revealed distinct stimulus‐locked and response‐locked activity (regression ERPs). A regression model was constructed that included stimulus‐locked activity, response‐locked activity, and an error term ε (a). Results showed no significant stimulus‐locked differences (b). However, a response‐locked CPP was observed (c). The gray bar indicates a significant temporal cluster.

2.2.2. Parametric Regression

Including note position p as a parametric regressor (Figure 4a) split stimulus‐locked activity into an intercept component (Figure 4b) and a parametric component (Figure 4c). The parametric results showed a significant stimulus‐locked cluster of activity centered at electrode FC1 (FC1, F3, FCz, C3, Cz, Fz) from 308 to 600 ms, p < 0.001, Cohen's d = 1.28. Our earlier response‐locked result was replicated at a significant cluster from −348 to 100 ms centered at O1 (O1, Oz), p = 0.0021, Cohen's d = 0.65. Similar results were obtained near these cluster centers when we excluded identified familiar trials: F4, 468–488 ms, p = 0.018, Cohen's d = 0.66 (stimulus‐locked) and O2, −116 to 100 ms, p = 0.005, Cohen's d = 0.68 (response‐locked). See Figure S5b,c for individual participants' scores.

FIGURE 4.

Parametric regression revealed an increasing stimulus‐locked positivity leading up to familiarity judgments. Notes preceding button presses were assigned a position p, which was then used as a parametric regressor in a regression model that included stimuli (notes), motor responses, and an error term ε (a). When the intercept term (b) and parametric term (c) were estimated, we noted a parametric effect that depended on note position and was maximal at over frontal electrodes (e). We also reproduced our earlier response‐locked result (d, f). The gray bars indicate significant temporal clusters.

To visualize the effect of note position on stimulus‐locked activity, we reconstructed the EEG using the rERPs. In other words, we combined the intercept rERP result with varying amounts of the parametric rERP result. The reconstruction shows how the stimulus‐locked response builds over time (becomes more positive) as the familiarity decision approaches (Figure 5).

FIGURE 5.

Reconstructed EEG shows the effect of note position on the stimulus‐locked signal. As note position draws closer to the button press (a), corresponding stimulus‐locked activity increases in amplitude (b). The gray bar indicates the significant stimulus‐locked temporal cluster identified in the parametric analysis.

Out of interest in the difference between unidentified familiar trials and identified familiar trials, we conducted an exploratory analysis in which we repeated our parametric analysis but modeled these trials separately. We excluded any participant with five or fewer trials in either condition and calculated the mean voltage for each participant and condition at the previously identified clusters (total N: 18). Besides being exploratory, this analysis had the additional caveat of low trial numbers, as discussed earlier—many participants had fewer than 20 trials in a condition. We then conducted paired‐samples t‐tests comparing identified familiar and unidentified familiar trials. We found no difference in the note‐position effect t(17) = 0.22, p = 0.83, Cohen's d = 0.05 and a small difference in the response‐locked effect, t(17) = 2.13, p = 0.048, Cohen's d = 0.50 (Figure 6).

FIGURE 6.

Exploratory parametric analysis, split by song identification. (a) Stimulus‐locked activity did not differ between unidentified and identified songs. However, there was a small effect of song identification on the response‐locked signal slopes and amplitudes (b). Splitting response‐locked activity by response time (fast, slow) revealed differences in slopes, but not amplitudes (c). The gray bar indicates a significant difference in amplitude, and the black bars indicate a significant difference in slope.

We then conducted a second exploratory analysis to examine the slopes of the response‐locked signals. In line with previous work (Kelly and O'Connell 2013; Murphy et al. 2015), we estimated the slope of the response‐locked rERPs in a 150‐ms moving window using MATLAB's polyfit function for each condition (unidentified, identified) and participant. This analysis revealed a difference in the response‐locked slopes—songs that could be identified were preceded by steeper ramping activity from −400 to −250 ms, t(17) = 2.45, p = 0.025, Cohen's d = 0.58. We also compared the mean response time for familiar unidentified songs and familiar identified songs. Response time was defined in two ways: from the start of song playback and from the onset of the note just prior to a familiarity judgment. Consistent with other CPP studies, steeper ramping was associated with faster response times (Kelly and O'Connell 2013; Murphy et al. 2015; Pisauro et al. 2017). In our case, identified songs elicited faster response times compared to unidentified songs. This held true whether we measured response time from the start of the song (p < 0.001, t(17) = 7.83, Cohen's d = 1.85) or from the final note (p < 0.001, t(17) = 7.07, Cohen's d = 1.67). See Table S1 for mean response times.

Finally, we repeated the above exploratory analyses to examine fast and slow familiarity decisions. Previous research suggests that the CPP is modulated by response time, with later ramping for fast responses and earlier ramping for slow responses (Kelly and O'Connell 2013; Murphy et al. 2015; Pisauro et al. 2017). Although the total number of “familiar” responses was lower than expected, we conducted an exploratory analysis where we split the response‐locked trials according to the median response time relative to the start of the song (Pisauro et al. 2017) before estimating the rERPs for each condition (fast, slow) and participant, as described in our main analysis. This revealed ramping signals for both fast and slow familiarity judgments (see Figure 6c). A comparison of the slopes revealed a small early difference from −1000 to −850 ms, t(25) = 2.38, p = 0.026, Cohen's d = 0.47. However, no difference in amplitude was observed using the same time window identified in our main analysis, t(25) = −0.30, p = 0.75, Cohen's d = −0.06.

3. Discussion

Our goal was to build on a small body of previous work showing neural evidence that recognition judgments rely on evidence accumulation. We expanded on this work in two ways. First, we used an unmixing procedure to isolate stimulus‐locked and response‐locked activity. This allowed us to examine the effect of “strength of evidence” on the stimulus‐locked signal independent of response‐locked activity. Second, we focused on familiarity judgments as opposed to recall judgments. As hypothesized, and in line with previous work (Evans and Wagenmakers 2019; Osth et al. 2018; Ratcliff 1978; van Vugt et al. 2019), our results revealed a prominent response‐locked accumulation signal. However, no such stimulus‐locked accumulation signal was observed—and no difference between familiar and unfamiliar stimulus‐locked responses—suggesting mixed evidence that song familiarity judgments rely on evidence accumulation.

In our first EEG analysis, we examined both stimulus‐locked and response‐locked activity for neural signals related to evidence accumulation. A prominent response‐locked CPP was observed, similar to response‐locked activity that has been seen previously in perceptual tasks and linked to evidence accumulation (Devine et al. 2019; Dou et al. 2024; Kelly and O'Connell 2013; Loughnane et al. 2018; O'Connell et al. 2012; Pisauro et al. 2017; Ruesseler et al. 2022; Twomey et al. 2015, 300). However, we found no evidence of stimulus‐locked evidence accumulation (i.e., no difference between familiar and unfamiliar notes). This contrasts with previous work showing a stimulus‐locked CPP during recall (van Vugt et al. 2019). We note two important differences between previous work and our work. First, our task emphasized familiarity, not recall. Second, we used an unmixing procedure to isolate stimulus‐locked and response‐locked activity. The reason why unmixing is necessary is to rule out the possibility that the observed activity is artefactual and the result of overlap between adjacent events (Frömer et al. 2024). Thus, our first EEG analysis suggests that familiarity decisions (but not familiar stimuli) elicit a CPP.

Our first EEG analysis examined average stimulus‐locked activity; our second analysis instead tested whether stimulus‐locked activity changes over time. By including note position as a parametric regressor we found evidence of a positive‐going change at frontal electrode sites: more positive for notes closer in time to the familiarity decision. This effect had a frontal scalp distribution that was distinct from the CPP observed at the time of the response, and distinct from previously observed central‐parietal stimulus‐locked activity (Kelly and O'Connell 2013; Murphy et al. 2015; O'Connell et al. 2012; O'Connell and Kelly 2021). However, frontal cortex is known to be involved in evidence accumulation during perceptual decisions (Brosnan et al. 2020; Rahnev et al. 2016) and frontal stimulus‐locked evidence accumulation signals have been observed during interval timing tasks (Ofir and Landau 2022) and perceptual judgment tasks (Gherman et al. 2024). Thus, our parametric result could be taken as evidence of a stimulus‐locked contribution to evidence accumulation. Caution interpreting this signal is warranted though due to its frontal distribution and due to alternative explanations for why a stimulus‐locked signal might change over time. For example, although we interpreted our parametric signal as an increasing positivity, it could also index a negativity that is decreasing over time due to adaptation (Wong et al. 2024).

The precise mechanism behind our stimulus‐locked effect is unclear and possibly beyond the scope of this project. Some have argued that frontal stimulus‐locked activity in the presence of a parietal evidence‐accumulation signal may simply reflect increased attention to decision‐relevant stimuli (Gherman et al. 2024). In line with this perspective, our stimulus‐locked effect had a timing and topography consistent with the P300, an ERP component linked with several cognitive processes including enhanced arousal to targets in a target detection task (Nieuwenhuis et al. 2005). Under this view, the attention‐related activity would be incidental and driven by an accumulation process nearing a decision threshold.

Others have argued that frontal signals such as error‐ and/or conflict‐related activity might contribute to the accumulation process, for example as an internal evidence signal (Desender et al. 2021; Murphy et al. 2015; Stone et al. 2022, 2024; Wendelken et al. 2009). In line with this perspective, the observed effects are consistent with an ERP component called the N400, which is related to the processing of meaningful stimuli (including tones: Kutas and Federmeier 2011). A key driver of the N400 and a key feature of music familiarity is the prediction of upcoming notes based on the preceding context (Leaver et al. 2009; Malekmohammadi et al. 2023). Although our task was not designed to elicit the N400 (i.e., we did not manipulate the notes), we speculate that note predictability—as indexed by the N400—could have served as internal evidence in favor of a familiarity decision. Put another way: A song might be judged as familiar if the notes start to feel predictable. The link between familiarity and prediction has been studied previously. For example, participants are more likely to report a feeling of prediction (i.e., of knowing what happens next) when viewing a familiar video compared to an unfamiliar video (Cleary et al. 2021). In general, predictions about the future rely on remembering the past (Schacter et al. 2007).

Our task was not designed to differentiate between “attention” and “prediction” explanations of our stimulus‐locked effect. Future work could manipulate attention using dual tasking or masking. By introducing unexpected (i.e., altered) notes, it might also be possible to examine the effect of stimulus predictability on familiarity decisions. These manipulations could help determine whether frontal activity contributes to evidence accumulation or is merely driven by it. This brings up another limitation of the present work. We are claiming here that song familiarity relies on evidence accumulation because we identified a response‐locked CPP. This is a reverse inference—inferring a specific cognitive process from an observed signal—the validity of which relies on the specificity of the observation (Poldrack 2006). Our confidence that the observed ramping activity is related to evidence accumulation as opposed to some other cognitive process comes from a careful reading of previous CPP work (Devine et al. 2019; Dou et al. 2024; Kelly and O'Connell 2013; Loughnane et al. 2018; O'Connell et al. 2012; Pisauro et al. 2017; Ruesseler et al. 2022; Twomey et al. 2015, 300) and the fact that we know of no other central‐parietal positive‐going decision‐locked signals. This previous work is high in internal validity but—we would argue—less ecologically valid compared to our task. Part of our motivation for the current work was to apply what has been learned previously about the CPP to a naturalistic task where measuring decision variables precisely is unfeasible.

Taken together, our EEG results suggest that song familiarity relies on evidence accumulation. This aligns with previous work in which evidence accumulation models have been applied successfully to recognition memory (Criss 2010; Ratcliff et al. 2004; Starns et al. 2012), including song recognition in flies (Pang et al. 2025). It also aligns with work in which these models have been applied specifically to familiarity (Cox and Shiffrin 2012, 2017; Gomilsek et al. 2025; Pan and Hu 2025; Weigard et al. 2024). Furthermore, our exploratory analysis suggests that familiarity plus recollection may elicit greater accumulation‐related activity compared to familiarity alone. This difference is not apparent in the stimulus‐locked signal, but rather around the time of the familiarity decision as an amplitude increase (and possible an increase in slope as well—see Figure 6b). Similar response‐locked differences are seen when comparing easy/difficult decisions in the domains of perception (Dou et al. 2024; Kelly and O'Connell 2013; O'Connell et al. 2012; Ruesseler et al. 2022), value (Pisauro et al. 2017), and memory (van Vugt et al. 2019). These differences have been interpreted as representing different rates of evidence accumulation (the “drift rate”)—steeper for “easy” decisions, shallower for “hard” decisions. In our case, this interpretation would suggest faster accumulation when familiarity is followed by recall than by familiarity alone. The enhanced accumulation signal, when viewed as “strength of familiarity”, is consistent with work showing that familiarity can improve source memory (Duarte et al. 2006; Hicks et al. 2002; Mollison and Curran 2012; Wais et al. 2008).

Many decisions are made gradually, not suddenly, and can be explained by carefully examining both behavior and neural activity in highly controlled laboratory experiments. We expanded this work to a naturalistic context and found neural evidence that familiarity decisions rely on evidence accumulation. Our work highlights both the challenge and the value of moving toward an ecologically valid understanding of the neural basis of decision‐making in memory and other domains.

Author Contributions

Cameron D. Hassall: conceptualization, formal analysis, investigation, data curation, writing – original draft, writing – review and editing, supervision. Jared R. Girard: investigation, writing – review and editing. Aaron Bishop: investigation, writing – review and editing.

Funding

This research was funded by a Natural Sciences and Engineering Research Council of Canada (NSERC) Discovery Grant to Cameron D. Hassall (Grant RGPIN 2024‐04848).

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Pooled participant response‐times (RTs) for familiar unidentified songs and familiar identified songs relative to start of song (a) and previous note (b). Dashed line indicates the mean RT for each condition.

Table S1: Mean response times for unidentified and identified songs.

Figure S2: Stimulus‐locked event‐related potentials (no overlap correction).

Figure S3: Response‐locked event‐related potential (no overlap correction). Shaded area and scalp topography mirror the same analysis window as the main analysis.

Figure S4: Design matrices for one participant's main analysis (a) and parametric analysis (b). Only the first 20,000 EEG samples are shown.

Figure S5: Participant rERP scores from the main manuscript. Error bars represent 95% confidence intervals.

Data Availability Statement

EEG dataset is available at https://doi.org/10.18112/openneuro.ds005876.v1.0.1. Analysis scripts are available at https://github.com/chassall/songfamiliarity.