37 Experimental Stimuli
(Modified on 2026-08-11)
2,099 words
In human-participant research, most inquiries involve exposing participants to visual or aural stimuli. In music research, this usually means audio or video recordings of music. The design and creation of musical stimuli is often the most important, challenging, and time consuming part of conducting human-participant music research. Don’t underestimate how long it will take to make stimuli, especially if you need to compose music or edit/manipulate audio by hand.
In human-participant music research, our main independent variables are often features of our stimuli. For example, our stimuli might consist of musical melodies, where the timbre, contour, mode, time-signature, cadence, or phrase structure are independent variables. In other cases, aspects of the stimuli might “only” be covariates that we need to control (or randomize). In experimental music research, we might directly manipulate such features, either by picking existing music which match the conditions we want, or creating new music that does.
The creation of musical stimuli necessarily involves a trade-off between ecological validity and control.
Ecological Stimuli
As we learned before, ecologically valid research can make claims that generalize to “natural” musical experiences. To achieve this, we can use “real” music—for example, commercial music recordings or public live performances—as stimuli. If we use “real” music in our experiments, it easier to make claims about “real” music outside the lab.
“Natural” music is very complex. Every piece of music, and every performance of music, is unique, with many independent attributes, all of which are potential research covariates. When we use ecologically valid stimuli, it is difficult to control for these variables. We can exert some control through systematic sampling of pieces—for example, quota sampling by genre, or tempo, or modality—, or through randomization. As always any aspect of stimuli selection/creation that can be randomized, should be. Whenever you use existing music as stimuli, try to sample pieces/recordings randomly from a relevant population of music.
Controlled Stimuli
In experimental research, it is especially important that we control covariates (and potential confounds) so that we can isolate the independent variables we want to test. To achieve more control over our stimuli, we can either manipulate existing music or generate stimuli entirely from scratch.
Semi-Artificial Stimuli
Let’s say we want to study the emotional impact of different modes, like Ionian, Dorian, Phrygian, Lydian, Mixolydian, and Aeolian. We could gather ecologically valid stimuli by collecting a stratified sample of pieces written in these modes—maybe 100 chants for each mode. We could exert more control by sampling only from one appropriate genre, like European medieval chants. Modes differ from each other in the scale steps they use—that’s what we want to study. But what if chants in different modes differ in other ways?
- Maybe chants written in Dorian tend to use more “arch-shaped” melodic contours?
- Maybe chants written in Lydian tend to use more ascending leaps?
- Maybe chants written in Mixolydian tend to be longer, but have more frequent cadential phrases?
These other potential covariates make it harder to make any claims about mode specifically. We could try to control for some of these covariates through sampling, like sampling only pieces that are the same length, use the same number of cadences and leaps, and which have the same overall contour. However, this would be difficult and might limit our sample size a lot. What’s more, there are probably other confounding covariates that we haven’t thought of.
So here is another approach: take all the chants in a given mode and recompose them into each of the other five modes. For example, take a Dorian mode and change all the F naturals to F#, so it is now (effectively) Mixolydian. We’d end up with, for each mode, 100 of the original chants as well as 500 recomposed chants. By comparing emotional responses between, for example, the 100 “real” Dorian chants with mode-altered Dorian chants, we’d be able to claim that the only difference between the two groups is the mode, not the contour, phrasing, or anything else. In general, taking existing music and systematically modifying it is a great way to create controlled, and yet still ecologically valid, stimuli.
Let’s consider another example. Perhaps we want to understand how listeners perceive phrase boundaries in Romantic solo piano music. To assure ecological validity, we might use real recordings of Romantic pieces. However, every performance of a Beethoven sonata or a Chopin etude is slightly (or dramatically) different. In particular, different pianists interpret and communicate phrases in the music in different ways: one performer might slow down a lot to emphasize every cadence (rubato), while another performer slows down less. Different recordings can also vary a lot in terms of sound quality, reverb, etc. Depending on our research goals, these performance variables themselves might interest us, or we might want to ignore them so we can study the pieces themselves (as notated). So how can we control these variables?
One approach is to randomly (or systematically) sample multiple performances of each of the pieces we want to study. We might even try to measure some parameters (like amount of rubato or reverb) and incorporate them into our analysis. Another approach is to generate fake performances using sequencing software to playback scores. By doing this, we can remove all performance details (rubato, dynamics, etc.) and thus isolate only the pitch and rhythm information (or whatever it is we want). Alternatively, we could systematically control performance details so that they are the same across our stimuli—for example, making sure that the use of rubato is always consistent. Using sequenced (usually MIDI) stimuli is common in experimental research, because we can use real pieces of music but at the same time precisely control the stimuli.
Note: just because you are using sequenced music is no excuse to make bad, unmusical stimuli. In the past, MIDI-generated stimuli often used cheap General MIDI patches, with perfectly quantized rhythm and no velocity changes—which sounds terrible. This might be appropriate for some research, but generally we can and should do better. With a little effort, you can make sequenced music sound very good, and still have total control over it.
Artificial Stimuli
To assure total control of our stimuli, the only option is to create the stimuli entirely from scratch. This allows us to create stimuli that only differ exactly in the way we want.
Much fundamental research in psychoacoustics was conducted using entirely artificial, synthesized stimuli. For example, Al Bregman’s (1990) essential work on auditory scene analysis was conducted using mostly sine-wave and noise generators. Check out this example from Bregman’s book:
See Bregman’s explanation of this from here.
Experiments using these sorts of stimuli—sometimes described as “beeps and boops”—were the basis for many fundamental discoveries about human psychoacoustics. For example, that humans can perceive frequencies from \(20_{Hz}\) to \(20_{kHz}\) and that the quietest sound humans can detect (\(0_{dB\ SPL}\)) has an amplitude of \(20_{\mu Pa}\). However, it is now well known that conclusions made using such artificial stimuli often fail to generalize to “complex signals” that occur in the real world.
Luckily, “artificial” stimuli doesn’t have to be weird, sinusoidal beeps and boops: We can compose and produce more realistic musical stimuli ourselves, or recruit composers/musicians to do it for us. The key is that we try to compose/produce/perform our stimuli in a systematic, controlled fashion. If, for example, we were studying responses to diatonic chord progressions, we could first generate a list of 12 different chord progressions that are all
- the same exact length;
- use the same harmonic rhythm, at the same tempo;
- use only the exact set of chords we want to study;
- are played with the same timbre and reverb; and
- are played in similar chord voicings, with similar voice leading.
We could then code these progressions into a MIDI sequencer or print a score and hire a pianist to play them for us to record.
In 2015, David Huron and I (2017) conducted an experiment to measure how the number of voices singing together (in a choral setting) affects the intelligibility of lyrics. To make stimuli for this experiment, we conducted a recording session with the Ohio State University’s (OSU) Chorale ensemble. The session was held inside OSU’s Weigel Auditorium, where we could capture the natural reverberation typical of a real choir concert (ecological validity), with microphones positioned to capture (as close as possible) this natural sound. In the session, we systematically recorded different subgroups of the choir (one, four, eight, sixteen, or twenty-seven singers) speaking and singing different words. In the single-voice condition, we randomly sampled equal numbers of soprano, alto, tenor, and bass singers to sing different words; For the larger groups, we systematically shuffled the subgroup of singers who sang each word. The words themselves were selected from a list of “phonetically balanced” English words, meaning that the phonemes in the words were a representative sample of phonemes in spoken English.
After collating and editing the results in a digital audio workstation (DAW), we had a total of 640 recorded stimuli, varying in length from 1.3–1.6 seconds. I had to manually export all 640 from my DAW, assigning a unique name to each one! Ultimately, through all of our efforts to control the process, we hoped that the only relevant differences in intelligibility between our 640 stimuli would be 1) the number of singers 2) whether the word was sung or spoken. All other variables (we hoped) were either controlled or “averaged out” through randomization.
General Considerations
Length
The length of stimuli is an important consideration. Stimuli in most experiments typically vary between 1–30 seconds. This allows ample time for a single participant to listen (and respond) to roughly 30–200 stimuli in a thirty-minute research session. Even when we use “real,” ecologically-valid music as stimuli, it is typically edited down to a short excerpt.
Generally, the longer an excerpt of music, the more complex and varied it is. This means that asking for simple self-report responses might make little sense: for example, if you ask a participant to listen to all 7+ minutes of Stairway to Heaven then ask them to rate the “intensity” of the music…what are they supposed to rate? A song like that spans the gamut of musical intensities! Will they even remember the beginning (recency bias)? Though Stairway to Heaven is an extreme example, you’ll find that even short musical pieces (like 1–2 minutes) can have a lot of variety in them. Using short stimuli allows you to focus participants’ attention on specific aspects of the music (variables), and makes it easier for them to hold the whole stimuli in their memory. It also allows you to get more responses from each participant.
Of course, the exact length of your stimuli will depend on your research question and the task you are giving to your participants. For questions concerning musical form or emotional impact, it may be necessary for a participant to hear a long passage or a whole piece in order to give a meaningful answer. Continuous-response self report, and some or other behavioral measurement protocols, may also lend themselves better to longer stimuli.
A final consideration is that you may, in some cases, want to control for length by making sure that all your stimuli are all the same length—either in terms of seconds, number of notes, number of chords, or whatever makes sense. In some cases, it can be hard to edit real music to make it an ideal length, but you might still try to get them all roughly the same length.
Loudness
One important musical factor that is very hard to control for is loudness. The loudness of stimuli can have a huge effect on how we respond to them—generally, louder stimuli is more arousing. Unfortunately, “loudness” is a complex psychoacoustic percept and it is surprisingly difficult to make stimuli “objectively” the same loudness; Acoustic measures like root-mean-square (RMS) only correlate approximately with perceived loudness. In many cases, you’ll have to simply rely on your own perception, or approximate “objective” approaches (like RMS), to match the loudness of stimuli.
Familiarity
Humans respond very differently to familiar music compared to unfamiliar music. Generally, when using “real” music as stimuli, the best practice is to avoid using music that your participants might recognize. This can make studying popular music difficult. Many researchers will ask participants to rate the familiarity of each piece they hear, so that this familiarity score can be referenced during their data analysis. In some cases, you might simply discard data points where a participant indicated they were familiar with the stimulus.