The Lucerne Groove Library: An audio stimulus corpus of 444 short Western popular music drum and bass patterns with behavioural, structural, and audio measurements

Previous corpus and expansion

We published the first version of the Lucerne Groove Library in tandem with an earlier study (Senn et al., 2018). The first version encompassed 248 drum patterns eight bars in length reconstructed from commercially available Western popular music recordings. The selection process for the patterns relied on a simple method consisting of first selecting 50 famous drummers and then five drum patterns played by each of these drummers, resulting in a selection of many iconic drum pattern stimuli. The current study’s expansion of the previous corpus consisted in two steps: firstly, we added bass lines to the drum patterns of the existing corpus. Secondly, we selected additional patterns based on the renown of the bassist and the famousness of the bass line and reconstructed both the drum pattern and the bass line. This resulted in a total of 444 combinations of drum patterns and bass lines.

Selection

In order to select the patterns based on the bass line, we first compiled a list of famous bass players. We collected names by asking professional bass players and bass educators, consulted online resources, magazines, and research articles. This resulted in a list of 246 bass players that we ordered by number of mentions in the consulted sources (between 1 and 22). We excluded bass players that mainly played jazz and double bass (e.g., Charles Mingus and Ray Brown), because their improvised walking lines follow different musical principles from the composed or at least semi-repetitive bass patterns that we focused on. Including double bass would also introduce a separate instrument with distinct timbral and articulatory properties, which could confound comparisons. We then selected the top 50 of the remaining bassists on the list, and five songs for each of these bass players. In several cases, the bassist played in a band with a drummer from the initial corpus, e.g., drummer John Bonham and bassist John Paul Jones of Led Zeppelin, or drummer David Garibaldi and bassist Rocco Prestia of Tower of Power. For these bass players, we judged how many of the bass lines in the original corpus were representative of their work and reduced the number of newly selected stimuli accordingly. However, we always chose at least two new stimuli specifically for the bass line. With these overlaps reducing the new material, we selected eight-bar passages from 198 songs for the expansion of the corpus that featured a famous, iconic, or simply style-typical and well-played bass line while avoiding bass solos. The consulted sources for player selection, online forums for bass players, and chart placement provided valuable help for selecting the songs, but sometimes the familiarity of the authors with a bass player’s repertoire led to a pragmatic choice. The selected passages had to be in 4/4 (or 12/8) meter, avoid playing techniques we cannot reconstruct, and be acoustically clear. Repetitive patterns (including variations) were preferred, though some less repetitive sections were also chosen. We selected the eight-bar segments from different formal sections—introductions, choruses, verses, interludes, (non-bass) solos—without systematically varying this parameter. Since these rules sometimes allowed for flexibility, the selection of songs and sections was influenced by the authors’ subjective views of what counts as well-played music or representative of a style, band, or individual player. In two cases, this led to two different sections of the same recording in the corpus, e.g., the Beatles’ “Come Together” for which we had the solo section in the previous corpus but wanted to include the iconic bass line in the corpus expansion. For each excerpt, we notated the song, artist, recording year, and location, and we time-stamped the start and end time of the selected passage in the version of the original recording we used (mostly obtained on iTunes). For each bass player and drummer, we also retained their full name, alias if applicable, birth year, and the country where they grew up.

Transcription and timing measurements

After the selection of the excerpts from the original recordings, we collected the necessary data for the reconstruction. As in the initial version of the corpus, we did this in two steps: (1) manual transcription of the music to identify and categorize all played events and (2) measuring the exact timing of each event. A professional drummer (author 3) transcribed the drums, and a professional bass player (author 2) transcribed the bass lines. Both transcriptions were checked afterwards by author 1, also a professional musician. The transcriptions included playing techniques (e.g., slap, plectrum, or harmonics for bass; rim clicks, fingers, or brushes for drums) and dynamics in four levels (1 = accent, 2 = loud, 3 = soft, 4 = ghost note). Some aspects were simplified in the transcriptions; for example, we did not measure and reproduce the exact pitch of the toms for each of the recordings. This kind of reduction limited the number of required drum instruments to 24. Some additional percussion instruments were included if they were deemed important for the drum pattern, even though they might have been performed by an additional percussionist (e.g., the shaker in Michael Jackson’s “Billie Jean”). After creating and checking the transcriptions, we measured the timing of each onset manually one by one using the Lucerne Audio Recording Analyzer software (LARA, version 2.6.3). Based on the transcriptions, all events were first marked roughly in a spectrogram and subsequently measured exactly in an oscillogram. Drums, due to their loud and percussive nature, are relatively easy to measure exactly. Cymbal onsets with their strong high-frequency content could easily be located using a high-pass filter on the audio. We are confident that the respective measurements are within a 5-ms error margin. For the bass, timing measurements proved to be more difficult: electric bass notes have a relatively long onset transient that makes it difficult to locate the exact beginning of a note, especially if other instruments play at the same time. For this reason, we (partly) separated the instrument stems with Izotope RX 8 (version 8.1.0.544) in some cases and took the measurements from an audio file that only included the bass (aside from some artefacts and other bass instruments). With this approach, we are confident that our bass onset timing measurements are within a 15-ms error margin. In total, we measured 85,941 events across the 444 clips (59,903 events for the drums, 26,038 the bass). For each of them, we stored the timing data (in seconds), the rhythmic position within the eight bars (in quarter notes), name of the instrument, dynamics level, and playing technique. For bass, we also stored pitch and rhythmic event duration. These two values were not measured (i.e., no measurement of note offset or average event pitch) but based on the transcriptions (durations were considered carefully during the transcription process).

Reconstruction

Based on the measured data, we created a MIDI file for each stimulus that contained all events and respective information in coded form (e.g., legato articulation and other playing techniques are triggered by silent MIDI events on specific pitches). As the bass samples had a long rise time, we applied a timing offset of −30 ms on all bass measurements for the reconstruction (−20 ms for slap playing style). The resulting MIDI files were imported into the Reaper digital audio workstation (version 6.14) for the reconstruction based on audio samples. In contrast to the first version of the corpus which features “neutral” audio drum samples, we decided to create five different, style-typical audio sample sets per instrument, i.e., five sets for bass and five for drums (Table 1). This allowed for the reconstruction of most of the patterns with sounds that are at least somewhat close to the original, thus improving ecological validity. The drum audio sample sets were created in Toontrack EZ Drummer 3 (version 3.0.6) and some library extensions, whereas the bass sample sets were created with plugins by Ample Sound (all version 3.1.0). Some audio sample libraries were used for multiple sets, but we used different instruments from these (e.g., different snare drums for pop, rock, and metal drums) or settings (e.g., different equalizers for funk and slap bass) for all sample sets. In addition to the five planned audio sample sets, we created one audio set for brushes and percussion.

For each drum sample set, we individually assigned the intended sounds to specific MIDI pitches. The MIDI mapping for each set of audio samples allowed us to deal with different perceived loudness of individual instruments or audio samples. The pre-sets and MIDI mapping for the audio sample sets can be found in the online repository. With their expertise for the respective instruments, authors 2 and 3 vouched for the quality and ecological validity of the resulting sound. In the next step, these authors assigned an audio sample set to each stimulus according to the similarity between the original sound and the sound of the audio sample set. In consequence, the audio sample sets have been used a different number of times across the entire stimulus set (Table 1).

Realization

The tracks featured two or three instruments: drums, bass, and, in 50 cases, additional percussion instruments. We decided not to apply panning to the different instruments in the mix (e.g., we did not assign them a specific location in the stereo space), as some future use cases might be affected by this. However, we varied the loudness of the tracks for each stimulus in order to better recreate the aesthetics of the original. We defined three different mix levels: standard (0.00 dB), low (− 2.00 dB), and very low (− 5.00 dB). With these levels, we could recreate the loudness discrepancies of the original instruments in the reconstructions; for example, when drums were much louder in the original, we put the bass track at − 5.00 dB. Afterwards, we applied the reverb track. As the drum and percussion samples had some built-in room sound and the bass samples were dry, the bass track had a higher send level on the reverb track (+ 3 dB) than drums (+ 1 dB). We used the Dragonfly Hall plugin (version 3.2.10) to create the reverb of a small-sized room with bright acoustics. The exact settings for the reverb can be found in the online repository. When this was completed, we rendered the tracks and saved the music to 24-bit WAV files. Lastly, we normalized the loudness of these files to − 20 LUFS.

Quality control

Authors 1–3 listened to all reconstructed stimuli to assess their quality. With 85,941 events measured by hand, errors were to be expected, and often these errors were clearly audible in form of rhythmic or dynamic inconsistencies. Each involved author either gave their approval of the reconstruction as an audibly truthful representation of the original, or flagged inconsistencies as precisely as possible. The respective section, timing, dynamics, or instrumentation was then inspected again in LARA, and measurements or associated data were corrected if found to be wrong. We proceeded with the newly generated MIDI file and reconstructed and realized the stimulus anew for another quality control by multiple authors until all authors approved the reconstruction. If the audio samples or the mix compared to the original recording was flagged as inconsistent, the respective settings were changed, and the result checked again by all authors to find the optimal sound and mix. Especially with respect to the timing measurements, it took several rounds for some stimuli to reach the final form, either because the audio was difficult to measure, or the music on the original recording contained rhythmic inconsistencies (which was verified visually by overlaying wave forms of original and reconstruction). We also dropped two stimuli that contained two drum sets playing different patterns simultaneously, because these conflicted with the technical limitations of our reconstruction method.

MIDI data

For each of the 444 audio stimuli, we derived 13 structural measures directly from the MIDI files, based on our timing measurements of the drum and bass events (Table 2).

Audio measurements

We measured 10 objective audio features of the 444 stimuli in the corpus with the MIR Toolbox (version 1.8.1, Lartillot et al., 2008) in MATLAB (version R2023b). For this analysis, we cut each audio file directly after the last onset for these measurements, because the presence or absence of long decays (e.g., in cymbals) affect the measurements. We selected measures that had been related to the urge to move or groove in previous studies (see Table 3). The exact commands for each measurement can be found in the online repository.

Participants

We recruited 917 participants (534 males, 379 females, 3 other) on the Prolific platform. They were 34 years old on average (sd = 11.598) and resided in South Africa (230), the UK (203), United States (109), Poland (69), Canada (30), Portugal (26), Spain (26), Germany (25), Greece (19), Hungary (17), India (16), Italy (16), Kenya (15), France (14), Netherlands (11), Australia (10), and 26 other countries (> 10). The majority self-identified as music listeners (742), a smaller proportion of the sample indicated being amateur (112) or professional (28) musicians, or music students (28). A small minority (7) were not interested in music. The sample of participants had a low Music Sophistication Index (MSI, Müllensiefen et al., 2014) Training score (mean = 18.050, sd = 10.164) compared to a UK standard population (Müllensiefen et al., 2014). Participants had a mean Goldsmith Dance Sophistication Index (DSI, Rose et al., 2022) Urge to Dance score of 24.134 (sd = 6.026), and generally liked popular music genres, with a mean rating of 4.040 (sd = 0.849) on a [0,6] Likert-type rating scale. The majority indicated having good English language skills (native speaker = 478, Common European Framework of Reference for Languages [CEFR] levels C = 344, B = 70, A = 25).

Measures

We measured six scales that are related to the psychological model of musical groove (Senn et al., 2019, 2023a) using self-report questionnaires. We measured participants’ experienced urge to move (MOV, 3 items) and their listening pleasure (PLE, 4 items) with the Experience of Groove Questionnaire (EGQ, Senn et al., 2020, 2024). We measured participant’s energetic arousal (ENE, 4 items), their rhythmic interest (INT, 3 items), and their inner representation of temporal regularity (REG, 4 items) with the questionnaires from Senn et al. (2023b). All items were answered on 7-point Likert agreement scales with range [0,6], and the respective item ratings were averaged to form the variables MOV, PLE, ENE, INT, and REG that can be analysed individually, in parts, or together in the SEM version of the groove model (Senn et al., 2023a). Participants also indicated their familiarity (FAM) with the stimuli by answering whether they had heard the music before (one item 5-point Likert scale: definitely not, probably not, I do not know, probably yes, definitely yes).

As in previous groove studies carried out by the authors, participants filled out the MSI Training scale (Müllensiefen et al., 2014) and the DSI Urge to Dance Scale (Rose et al., 2022). Participants also indicated their liking for 21 music styles and assigned each stimulus to one or more musical styles (optional task). We used a weighted calculation to assign each stimulus a style that is based on observed and expected style assignments (for details, see the supplementary of Senn et al., 2021).

Procedure

We followed the procedures outlined in Senn et al. (2024) and Bechtold et al. (2025c), but we adapted the randomization scheme such that all stimuli were rated by approximately the same number of participants and always appeared in different combinations. We set up the questionnaire on the SosciSurvey platform (www.soscisurvey.de). The participants were informed about the study and gave informed consent by mouse click. Participants answered the demographic questions and indicated their music preferences. Then they heard an example to adjust the playback volume to a comfortable level. Subsequently, a randomizer selected 10 stimuli, and the six main blocks of the experiment began. In the first block, the participants heard the 10 stimuli (one per survey page), indicated how familiar they were with the music (FAM), and optionally assigned the music to a style. In the second block, participants heard the same stimuli again, but rated the urge to move (MOV) they experienced in response to the music (3 items). In blocks three to six, participants heard the stimuli again and filled out the questionnaires for pleasure (PLE), time-related interest (INT), energetic arousal (ENE), and inner representation of temporal regularity (REG). Which scale appeared in which block was randomised. Between blocks, participants filled the Goldsmith Music Sophistication Index Training Scale and the Goldsmith Dance Sophistication Urge to Dance Scale. After the sixth block, participants received some information about the study. On average, participants needed 36 min to complete the study. They were remunerated with GBP 6.

Statistical analysis

We primarily conducted descriptive analysis across several measures to characterise the corpus. We used R (version 4.4.2) in the RStudio environment (version 2024.09.1) and created all figures with the ggplot2 package (Wickham, 2016; version 3.5.1). A more thorough analysis of the data within the groove model framework (Senn et al., 2023a) will be published later in a separate article.

Styles

Participants optionally assigned each stimulus to one of 21 styles. Based on these answers, every stimulus was assigned to one style (Table 5); 14 styles were assigned at least one stimulus. Rock, funk, pop, jazz, and rock ‘n’ roll form the largest groups, but all styles except rap/hip-hop encompass sufficient stimuli with a variety of urge to move ratings for future selections based on style.

MIDI data

We calculated 13 variables based on the MIDI data of the stimuli. Table 6 gives descriptive data of these measures and their correlation with the urge to move scale.

For eighth note swing, 404 stimuli featured no swing, and only 48 had swing eighth notes (ratio > 1.5). For two stimuli, eighth note swing is not given (NA), because they featured less than eight eighth notes. Forty-two stimuli featured sixteenth note swing (ratio > 1.5), 264 did not; for 138 stimuli, no sixteenth note swing is given (NA), because they featured less than eight sixteenth notes. Microtiming can be grouped for example as follows: we consider 55 stimuli to have loose timing (> 0.02 s), 347 have average microtiming (between 0.01 s and 0.02 s), and 42 stimuli are tight (< 0.01 s). For the mean subdivision, a total of 234 stimuli are based on eighth notes and 184 feature mostly sixteenth notes; 22 are based on eighth note triplets, and four on sixteenth note triplets. The distributions of the MIDI data-derived variables, either numerically as in Table 6 or in groups as above, allows researchers to select stimuli for experiments based on a variety of measures. The corpus includes extremes as well, e.g., very fast and slow tempos, short and long stimuli, as well as rhythmically dense stimuli and stimuli with little rhythmic activity.

Audio measurements

We provide 11 MIR measurements based on the audio of the reconstructed stimuli. As Table 7 indicates, most measures show some variety, and several of these are also correlated with the urge to move scale (MOV). RMS loudness has very little variability because we normalized the loudness of the stimuli to facilitate their use in listening experiments.

Contextual data

The stimuli cover an extended period of popular music history: 1956–2017 (mean = 1983, sd = 13.7 years). Two stimuli have no associated recording year, as the album was recorded over multiple years. Figure 2 shows the distribution of stimuli over half-decades. Stimuli between 1965 and 1985 are strongly represented. For the earliest (1955–1965) and latest (2015–2020) years, we selected only a few songs. The recording year is not correlated with the urge to move (p = 0.535) on stimulus level.

The corpus features stimuli as played by 117 different drummers and 119 bass players. The musicians who were featured most often are drummer Ringo Starr (9 stimuli) and bassist Pino Palladino (11). The corpus represents the music of 177 acts/bands, with highest frequencies for James Brown (10) and The Beatles (9). The extracts were from 350 albums, several of which are featured multiple times. Jimi Hendrix’s album “Are You Experienced?” and the Pixies’ “Doolittle” topped this chart with four appearances each.