Soundtrap usage during COVID-19: A machine-learning approach to assess the effects of the pandemic on online music learning

Abstract

The COVID-19 pandemic prompted a sudden rethinking of how music was taught and learned. Prior to the pandemic, the web-based digital audio workstation Soundtrap emerged as a leading platform for creating music online. The present study examined the growth of Soundtrap’s usage during the COVID-19 pandemic. Using machine-learning methods, we analyzed anonymized user data from Soundtrap’s 1.6 million educational users in the United States to see if the pandemic affected Soundtrap’s education user base and, if so, to what extent. An exploratory data analysis demonstrated a large increase in Soundtrap’s user base beyond five standard deviations beginning in March 2020. A subsequent changepoint analysis identified March 17, 2020, as the day this shift occurred. Finally, we created a SARIMAX model using data prior to March 17 to forecast expected growth. This model was unable to account for user growth after March 17, showing highly anomalous growth rates outside of the model’s confidence interval. We discuss how this shift affects music education practices and what it portends for our field. In addition, we explore the role of machine learning and artificial intelligence as a method for research in the music education field.

Introduction: Digital audio workstations in music education

School music in the United States follows a long and storied tradition of synchronous, in-person music teaching and learning (Mark & Gary, 2007). In broad terms, this usually involves lessons in elementary schools following an approach modeled on the teachings of Kodály, Dalcroze, or Orff, and large-ensemble instruction in middle and high schools (Benedict, 2010; Elpus & Abril, 2019). For logistical, technological, and economic reasons, having one teacher work with many students simultaneously in a physical space has been an obvious and necessary paradigm for music instruction.

A major growth area for music education in recent years has been at the intersections of music technology and popular music, often including moves away from the dominant performance-oriented paradigm toward practices centered on making original music with digital tools. As part of this change, the use of digital audio workstations (DAWs) to allow students to create, arrange, and store music online has increased in music education classrooms (bell, 2015a). Some of today’s most popular DAWs include GarageBand, Pro Tools, Ableton Live, and Soundtrap. Most DAWs show visual representations of recorded audio from left to right, with layered “tracks” (individual audio files) shown from top to bottom. In addition, many DAWs offer users a library of samples and loops that they can add to their compositions. These include beats (i.e., percussive tracks mimicking or sampling drums that often include bass lines and synthesizer sounds) and a range of short melodies and harmonic progressions on a wide variety of virtual instruments. Within the DAW, users can alter musical parameters such as pitch, key, tempo, equalization, panning, and dynamics. DAWs also often provide a range of effects (e.g., chorus, flange, tremolo, reverb, echo, delay) alongside a selection of simulated specific microphones, amplifiers, guitars, and keyboard sounds.

As technology has become more affordable, and DAWs such as GarageBand come free on Apple computers and iPhones, the incorporation of DAWs continues to expand in music education settings (Clauhs, 2020). This presents an exciting prospect for music educators as many learners of all ages and abilities have access to an all-in-one musical-instrument-recording-studio at all times (bell, 2015a). At this point, many students in K–12 classrooms can be considered digital natives (Prensky, 2001; Thomas, 2011) and many are increasingly becoming familiar with audio recording and editing through digital applications such as TikTok. This being said, of course, not every young person confronted with a digital musicking interface will be excited by the creative potential it offers; nonetheless, DAWs present new opportunities that prior technologies did not.

Facilitating musical experiences with students who grew up using technology provides opportunities for music educators to shift the focus from music education as performance to music education as computer-mediated creativity through media (Thibeault, 2014). This shift might also bring more students into to the music education classroom, as DAWs can “appeal to a wide range of students, many of whom might not participate in music-making otherwise” (Walzer, 2016, p. 24). In addition, the incorporation of DAWs into the music classroom can allow music educators to facilitate opportunities for student composition (Ruthmann, 2007) and other creative projects (Rosen et al., 2013).

Soundtrap: Web-based DAW

The COVID-19 pandemic prompted a sudden rethinking of music teaching and learning. Portable computer technology, widespread internet access, and online music-making software enabled a previously unthinkable shift in music education praxis. When the pandemic began in early 2020 and school districts across the country closed and local and state governments issued “lockdown” or “rest-in-place” orders, schoolteachers were suddenly thrust into an unknown realm (Smith, 2020). The notion of asynchronous instruction was anathema to most music teachers as recently as the beginning of 2020. But by mid-March of that year, the coronavirus pandemic had forced everyone out of their classrooms, and technology made teaching and learning online a viable proposition, and in some instances, mandatory. While research and pockets of practice pointed to the potential for engaging students via technology and especially the internet, many teachers lacked familiarity with such modes of instruction (Cayari, 2021). Some music educators found themselves in unfamiliar territory: reliance on technology; project-based learning; informal pedagogies; emphasis on play, composition, and improvisation; peer learning; and a lack of direct, synchronous instruction.

Many teachers turned to various internet-based music-making and learning platforms, including BandLab, Acapella, Auralia, Noteflight, and others to serve a range of creative and other learning goals. Among these platforms was the popular DAW Soundtrap. Soundtrap is a cloud-based DAW that enables musicians to collaborate synchronously and asynchronously to make music, including the option to communicate via in-app video conferencing. There are consumer and educational versions of the software. Educational accounts may be purchased in bulk by schools and school districts, and provide a “walled garden,” a secure site with no access to the public or anyone outside of that subscription group. Soundtrap also offers free training courses to its users. Throughout the spring and summer of 2020, these training options included Soundtrap “Expert” and “Educator” certifications, which provided customers with an introduction to the functionality of the DAW and, in the case of the training designed for educators, suggestions on how to use it for music learning.

One appealing aspect of Soundtrap, and some other DAWs like BandLab, is that it works on a range of hardware devices, including Mac and PC computers, tablets, and smartphones. This arguably made these platforms more accessible during lockdown periods of the pandemic, when some students did not have access to a computer or tablet. In addition, Soundtrap is cloud-based; students can share projects and collaborate using an internet browser without the need to download and transfer projects. This collaborative aspect of cloud-based DAWs allows students to work on projects simultaneously, instead of one at a time; an analog in the word-processing domain is Google Docs. While any level of technological requirement means that a platform or instrument is not going to be accessible to every potential user (bell, 2015b), the relative accessibility of Soundtrap made it a viable option for schools and teachers during the pandemic in ways that other existing platforms were not.

Soundtrap, like other DAWs, is designed to be used for purposes of music creation or composition. It is necessary to construe “composition” in a way that accounts for contemporary, 21st-century modes of music creation. For instance, working in a DAW no longer requires reliance on older modes of music notation, such as the (paper or electronic) five-line staff. In the DAW context, this is substituted by the visual representation of audio waveforms and MIDI clips that also offer greater utility within the software. The roles of “composers,” too, in the realm of the DAW to include arranger, performer, improviser, engineer, technician, songwriter, and producer (Moir & Medbøe, 2015).

The present study examined the growth of Soundtrap’s educational user base during the COVID-19 pandemic. Specifically, our research aimed to determine (a) if there was a growth in Soundtrap’s educational user base during the pandemic; (b) if so, the size of this growth; and (c) if this growth was an anomaly, unable to be predicted by pre-COVID new user rates. Although commonsense interpretations of educators’ responses to the pandemic indicate an increase in the use of Soundtrap during this time, it is nonetheless helpful as a matter of historical record to verify and precisely document this increase. This documentation may provide a foundation for future research about the use of online platforms in music education as well as the longer term effects of the pandemic. Researchers were given access to anonymized user data to determine if the pandemic affected user rates on the platform, and, if so, the scale of these effects. Although this study only examined data from one platform, Soundtrap is among the most widely adopted platforms for music educators. As such, we believe the results speak more broadly for the music education field. Understanding the extent of Soundtrap adoption during the pandemic might serve as a proxy for music educators’ overall engagement with DAWs and increased reliance on technological resources during the height of the pandemic. For this research study, we used methods that we believe to be new in music education research literature; the machine-learning methods we describe in the following sections appear to be a first in music education scholarship.

Method

Data were obtained from Soundtrap, which enumerated every user on the platform. This included more than 10 million users beginning from the platform’s inception on January 1, 2011, to January 31, 2021, when data were obtained. User information received from Soundtrap was anonymized, and the only variables available for analysis were: the date created, country, and account type. Account type refers to Soundtrap’s different user types of “consumer” and “educational.” We filtered this data set according to our research questions, and only examined the educational users in the United States. Although these accounts began from January 17, 2014, we further filtered the set by date, only examining the accounts from January 31, 2015, and onward. We did this because, according to Soundtrap, these first few educational accounts were often test accounts created internally before launching their educational version of the platform. In addition, by beginning our analysis from this date, the data set was trimmed to exactly 6 years, making the seasonality of the time series more regular. This final data set included approximately 1.6 million users.

Because of the size and scope of data used in this analysis, more familiar parametric tools were unavailable. Instead, the researchers chose to interrogate the data using machine-learning methods most often associated with big data analysis. Although lacking a precise agreed-upon definition, most data scientists agree that the term “big data” describes large volumes of data that may be examined for patterns using computationally sophisticated tools (McAfee et al., 2012). We believe the present study is without precedent in the music education field in terms of the size of the data analyzed and the kinds of machine-learning methods employed to do so. However, analysis of big data is commonly used in related fields, including a subfield of musicology termed “computational musicology” (Bel & Vecchione, 1993), music theory (Anders, 2021), composition (Briot & Pachet, 2020), and digital humanities (Burgoyne et al., 2015). The data in this study were not exhaustively large by big data standards, but the size of the data in this project was too large, and the analytical tasks below were too computationally expensive for consumer computers. Thus, we used a big data approach for our methodology. The data were ingested into a MySQL database running on a Ubuntu server. A separate Windows server running R Studio was given access to the server and used to query the data and create data objects within R to be analyzed.

R is an open-source research language and environment designed for statistical computing and formatting of graphical outputs (Ihaka & Gentleman, 1996). More common in fields where large data sets are the norm, like economics and the physical sciences, R allows users to manipulate large amounts of data, compute a wide range of statistical variables for analysis, and format graphics from these analyses to tell a visual story of the data. R was selected for this project because of its ability to handle large amounts of data, its large user base of support, and its long history in data science.

Once in R, the data were analyzed using three layers of analysis. The first layer was an exploratory data analysis (Myatt, 2007) that allowed the researchers to visually inspect the time-series data for possible trends or seasonality. Following this inspection, a changepoint test (Hinkley, 1970) was used to determine any precise changes in means during the time series. These two layers of analysis allowed us to answer our first two research questions, specifically to determine if there was a change in Soundtrap’s user base, and, if so, the size of this change. Finally, a Seasonal Autoregressive Integrated Moving Average (SARIMA) model (Brockwell & Davis, 2016) was used to determine if this change in Soundtrap’s user base was anomalous and could not have been predicted by prior new user rates. These methods of analysis are discussed in more detail in the following section.

The authors sought and were granted institutional review board approval for the present study from each of our respective institutions. Due to the data being obtained by a third party, and because those data were anonymized with no personally identifiable information, this study was considered exempt from human subjects review. There is a possibility of a conflict of interest in the present study due to the fact that the analyses below depend upon the authors being provided access to Soundtrap’s proprietary data. It is possible that we would only be given access to data that are in the interest of the corporate entity holding the data. It is also possible that, as authors who wish to have continued access to proprietary data, we might conduct research that encourages this access. Although research that involves the participation of a non-neutral third party may raise valid concerns, we would argue that the benefits of analyzing a data set of this size and scope—and one that is only available through a third party—outweigh possible conflicts of interest, especially given the particular significance of the COVID-19 pandemic as a unique context for music education. In addition, there have been no limitations placed on the researchers by Soundtrap, other than ensuring that our reporting of the data does not violate securities and exchange regulations in the United States that prevent the disclosure of internal company information that shareholders do not have access to. In our analysis below, explicit daily new user rates have been removed upon request from Spotify—Soundtrap’s parent company—and are instead displayed using arbitrary units. This allowed the researchers to conduct our analyses without violating these regulations. Finally, there is no financial relationship between the researchers and Soundtrap, other than some of the researchers being users on the Soundtrap platform.

Results

This study investigated the use of Soundtrap during the COVID-19 pandemic by examining educational account data in the years before and during the onset of the pandemic. This examination consisted of three layers of analysis, beginning with an exploratory data analysis (Myatt, 2007), followed by a changepoint analysis (Hinkley, 1970), and finally developing and testing a predictive model using a SARIMA model (Brockwell & Davis, 2016).

Exploratory data analysis

For the initial exploratory data analysis, a plot of daily new users was generated for educational users in the United States. This time series spans from January 31, 2015, to January 31, 2021, and uses a 7-day moving average (MA). Visual analysis of the daily new users plot (Figure 1(a)) shows a regular time series from January 31, 2015, through March 2020. This first segment demonstrates consistent trend and seasonality corresponding to school semesters. The trend is seemingly interrupted in March, when there is a sudden change in mean, while the semester seasonality persists.

Figure 1.

(a) Number of Daily New Educational Users for Soundtrap in the United States. (b) Differenced and Scaled Daily New Educational Users for Soundtrap in the United States.

Before determining the precise changepoint, and to further explore the data, we examined the data for stationarity. A Llung–Box test demonstrated significant nonstationarity (χ² = 60,633, p < .001). With the data determined to be nonstationary, we differenced the time series to produce a stationary plot. Differencing a time series removes any trend and seasonal effects, creating a time series that shows the variances of consecutive values. This demonstrated the extent that daily values deviated from a stationary (i.e., detrended and deseasoned) plot. To demonstrate the distance of each data point from the mean, this differenced time series was normalized and displayed using standard deviations. This normalized time series (Figure 1(b)) demonstrated the extent that data beginning around March 2020 deviated from normal, at times exceeding five standard deviations.

Changepoint detection

Although our visual analysis demonstrated a high level of confidence that a changepoint occurred in March 2020, we wanted to select a precise changepoint for our SARIMA modeling. We selected a changepoint detection method that could detect multiple changepoints, if any were present. The Pruned Exact Linear Time (PELT) method was developed to detect changepoints in large data sets, like genomes, where accurate detection is desired without a loss in computational time (Killick et al., 2012). The PELT method tests for various combinations of changepoints, referred to as segmentations, looking for the combination with the least cost function. Research into the accuracy of the PELT method has demonstrated that it is comparable with the At Most One Change (AMOC) method for single changepoint detection and Binary Segmentation for multiple changepoint locations (Killick & Eckley, 2014; van den Burg & Williams, 2020).

With our data, the PELT method produced a matrix of 19 segmentation combinations that included 30 possible total changepoints. Every segmentation had changepoint clusters in March 2020. Because the minimum segment length in our PELT test was set to 1 week, March 16, March 23, and March 30, 2020, were the most commonly identified changepoints. Of the 18 most conservative segmentations, there were no changepoints identified prior to March 16, 2020. This indicated a dramatic shift in mean occurring sometime in middle to late March. The most conservative segmentation identified only one changepoint on March 23, 2020 (Figure 2(a)).

Figure 2.

(a) Mean Segments of Daily New Educational Users for Soundtrap in the United States, With One Changepoint Identified. (b) Mean Segments of Daily New Educational Users for Soundtrap in the United States, With Three Changepoints Identified.

The second most conservative segmentation included three changepoints, at March 23, May 15, and August 24 (Figure 2(b)). However, from this plot, it is clear that these additional changepoints are correlated to seasonality, with a dip during the summer months. Because the function of differencing expects annual trends to be highly correlated, the scale of the anomalous findings after March 23, 2020, causes additional changepoints at the beginning and end of the semesters to be identified by the PELT test. Therefore, we selected the single changepoint segmentation as the most appropriate representation of change in mean.

To confirm this choice and further pinpoint an accurate changepoint, we ran an additional changepoint test using the AMOC method. Research has indicated this method is slightly more accurate than PELT, if it has already been determined that there is a single changepoint (van den Burg & Williams, 2020). Results of the changepoint test using the AMOC method demonstrated a changepoint on March 17, 2020. This result corroborated the results of the PELT test and reflects the experiences of many students and teachers who were suddenly moved to online learning around this time (Smith, 2020).

Forecast model

With the changepoint identified as March 17, 2020, we created a forecast model based on the first time-series segment from January 31, 2016, until March 17, 2020. These pre-COVID training data were used to create a predictive model of new daily users to the platform. Using this model, we were able to forecast what new user rates might have been after March 17, 2020, and determine how anomalous the daily user rates during COVID were.

For our forecast, we selected the Seasonal Autoregressive Integrated Moving Average with exogenous factors (SARIMAX) model. Like all ARMA models, SARIMAX consists of an autoregressive (AR) and an MA component. The former regresses the model based on its prior lagged values, and the latter incorporates prior error terms to create an MA. Figuratively speaking, the AR component creates a short-term memory based on immediate past results, while the MA component creates a longer-term memory based on the accumulated errors of past results. Put together, ARMA models are useful in forecasting future results based on prior events. ARMA models are commonly used forecasting models in economics and physical sciences and have recently been used to investigate the effects of the COVID-19 pandemic (Auger et al., 2020).

The SARIMAX model in the present study has three elements added to the simpler ARMA model. The first, which is shared by all ARIMA models, is an integrated function (I) that incorporates a differencing step for nonstationary time series. This was necessary to account for our time series’ nonstationarity. In addition, SARIMAX also incorporates a seasonal function (S) by reproducing the three AR, I, and MA functions, but according to the time series’ seasonal element.

In the present study, we began by using SARIMA in our forecast. However, because of the complex seasonality of the data as evidenced by the peaks and troughs correlated to academic semesters, we decided to instead use a SARIMAX model for greater accuracy (Arunraj & Ahrens, 2015; Manigandan et al., 2021; Vagropoulos et al., 2016). SARIMAX models allow for the inclusion of additional exogenous elements (X) into the model. These elements may be any other function or variable that increases model accuracy. In our SARIMAX model, we created a Fourier function consisting of two sine and cosine pairs, one with a once-per-year cycle, and another with a twice-per-year cycle corresponding with academic semesters. Visual and statistical analysis demonstrated that adding this Fourier function created a model that was more accurate to the pre-COVID data than the SARIMA model alone. Graphs of the SARIMAX model’s residuals and residual distribution from the pre-COVID training data demonstrated Gaussian residuals that were normally distributed around zero and with minimal skew (Figure 3).

Figure 3.

Left: SARIMAX Forecast Residuals Versus Time From the Pre-COVID Training Data in Arbitrary Units. Right: Projection of the Residuals Integrated Over All Training Data in Arbitrary Units.

Note. The (black) curve shows a fit to the data using a Gaussian function to demonstrate normally distributed residuals.

With the model created, we turned to examine the ability of the model to predict new users during COVID. Visual analysis already demonstrated a large shift in March 2020, with new users during COVID exceeding five standard deviations from normal. In addition, we detected a change in mean on March 17, 2020.

Although our model was successful in accurately describing new users in the pre-COVID training data, we expected it to perform poorly with the COVID test data. Results of the forecast demonstrated the SARIMAX model (2,1,2) (1,0,1,12) was unable to accurately account for the number of new users during COVID, with the rate of new users far exceeding the model’s confidence interval of 95% (Figure 4).

Figure 4.

Daily New Educational Users for Soundtrap in the United States.

Note. The (gray) plus points show the pre-COVID data used to train the SARIMA model, and the (black) diamond points show the COVID test data. The (red) solid line shows the SARIMAX model performance after training and the (blue) dashed line shows the SARIMAX forecast. The blue band shows the 95% confidence interval on the forecast model.

Summary

Exploratory data analysis showed a sudden uptick in Soundtrap users in March 2020. When detrended, deseasoned, and normalized, this increase represented a shift that was at times more than five standard deviations above normal. Put simply, Soundtrap’s increase in users beginning in March 2020 was highly abnormal. To confirm if and when a change had occurred, we conducted a changepoint analysis. Results from the PELT changepoint method demonstrated a cluster of segmentations that included dates in the middle of March. An additional AMOC examination identified March 17, 2020, as the most likely single changepoint.

To further examine how anomalous data after March 17, 2020, were, we created a SARIMAX forecasting model to predict what daily new user rates would have been after March 17, 2020, using the previous time-series segment as our training data. Using a SARIMAX model (2,1,2) (1,0,1,12), we included an exogenous Fourier function to more accurately simulate the time series’ complex seasonality based on academic semesters. The model performed very well against the pre-COVID training data, but exceptionally poorly using the COVID test data.

Across our three phases of analysis, there was overwhelming evidence that a sudden and drastic shift occurred within Soundtrap’s educational user base in the United States around March 17, 2020. This shift occurred alongside sudden and drastic changes to K–12 education in the United States following the beginning of the COVID-19 pandemic. Although causation cannot be ascribed with the present study’s descriptive design, the authors are confident in attributing this increase to the COVID-19 pandemic based on a facial claim to validity.

Implications for music education

The data collected and analyzed in this study show the rate at which new Soundtrap educational accounts were created before and during the COVID-19 pandemic. As discussed previously, we believe it is important to verify and measure the extent of this increase for historical record. While DAWs have become more commonplace in school music programs over the past decade (bell, 2015a), the drastic increase in the use of DAWs at the start of the COVID-19 pandemic presents interesting challenges and possibilities for the music education field, especially to those working in preservice music teacher education and creators of music curriculum and content standards. At the start of the pandemic, DAWs moved from being used primarily in specialist music technology and production classes, to more widespread adoption by music teachers of all content area specialties, including large ensembles and general music (Cayari, 2021).

This research team did not have access to data demonstrating exactly what activities teachers and students undertook with their Soundtrap accounts (e.g., songwriting, composition, recording vocal or instrumental parts for virtual ensembles). Similarly, the data do not provide insights into sustained curricular integration of Soundtrap by teachers or continued engagement on the part of students after the initial influx of account creations in 2020. There is a need for future research to determine the extent to which music teachers and their students continued to use Soundtrap once in-person instruction resumed. Should future research show that there has been continued engagement with Soundtrap and other DAWs coming out of the pandemic, it would strengthen the case for the incorporation of more DAW-based training in preservice music teacher education (Bauer & Dammers, 2016).

Another implication of this research for music educators is the need to explore equity and access when it comes to the use of DAWs. Although DAWs have become increasingly user-friendly, and Soundtrap is accessible via a range of hardware devices, including tablets and smartphones, utilizing the platform requires a level of access to technology and funding not available to all students. This is especially salient, for instance, in rural or urban schools that tend to be less well funded in comparison with suburban schools in the United States, and in cases where students do not have access to resources of time or technology at home (Catalano et al., 2021). The School District of Philadelphia, for example, banned teachers from requiring students to work remotely or grading virtual participation during the remote-instruction period of the pandemic as not all students had equal access to technology (Mezzacappa & Wolfman-Arent, 2020). While DAWs and other digital technological solutions offer exciting opportunities that can be democratizing to an extent, it is important to bear in mind local and individual circumstances that prevent technologies from providing a panacea.

This study demonstrated that, during and following the transition to internet-based music teaching and learning during the COVID-19 pandemic, many music educators turned to Soundtrap to assist them in facilitating music learning activities. Time will tell how much of this creative and technology-based activity will carry over in the years to come, now that, as of fall 2022, almost all schools have returned to in-person instruction. Will the affordances provided by DAWs such as Soundtrap be compelling enough to attract music educators who might seek to return to “normal” in their music programs? As more user data become available from Soundtrap, it will become clearer whether the COVID-19 pandemic was a tipping point for the use of DAWs in music education or merely a life raft until such a time as music educators could return to business as usual.

Emerging music education research

As digital platforms such as Soundtrap proliferate and come to occupy greater space in society, data collected by them are able to help researchers answer previously unanswerable questions (Rudder, 2014). The present study incorporated machine-learning methods that are novel to music education, a research field typified by educational psychology and its contingent methods (Colwell & Webster, 2011; Yarbrough, 2002). As the educational sphere—including music education—becomes more digitized, there is an opportunity to ask new questions of the abundance of available data, or to ask old questions in new ways, leading to new understanding. As an example, previous research regarding musical preferences has been conducted using smaller scale studies observing students’ behaviors (LeBlanc, 1979). Online platforms like Soundtrap, on the contrary, are able to offer data on students’ preferences at a scale in the millions. By using sophisticated machine-learning models, the relationships between these vast data can be further examined for complex patterns.

We can also ask questions regarding students’ creative behaviors and processes using data collection methods that were previously unfathomable. Whereas these questions would have previously required smaller scale-observations of in-person learning episodes, now they can be asked of an enormous swath of music learners across age groups and geography, without the limitations of an observer inside a single classroom. Because platforms like Soundtrap are both the instruments for learning and contain the products and processes of this learning, students’ creative behaviors are automatically encoded as data. The creative output is entirely digitized, as is the user interface and process that led to students’ creative output. The limitation, then, is not the availability of data, but the creation of analytical methods that appropriately use this wealth of data to draw valid and useful conclusions for music learning. As practices and practitioners in our field move further into digital spaces, we will need to develop new research methodologies that respond to these new spaces and attend to the different ways music-making and learning are taking place. One possible response to this opportunity, as the authorship of this article demonstrates, is a collaboration between music education researchers and data analysts, data scientists, and individuals in other allied fields to ensure the data are being appropriately interrogated.

Finally, while the present study uses machine learning to describe a phenomenon within music education, machine learning and artificial intelligence are increasingly being used as creative tools within music learning. Soundtrap, for example, integrates a recommendation system for selecting loops based on the elements a student has already added to their project. Rowe et al. (2016) describe an improvisation companion for elementary-aged students that can learn their musical vernacular and create novel responses based on a student’s musical call. The musicians and scientists Carré and Pachet have developed what they call a flow machine, which is able to translate musical elements from one piece or body of music into another. This allows users to, for example, autonomously create a novel melody based on the work of The Beatles that is bespoke to a user’s chord progression, rhythm, accompaniment, or some other musical parameter (Avdeeff, 2019; Pachet et al., 2021). As these artificial intelligence tools become more integrated within music learning tools, our field will need to begin considering their educational purposes and effects.

Author biographies

David H Knapp is an Assistant Professor of Music Education at Syracuse University, where he teaches modern band, music technology, steel band, and philosophy of music education. His research has been published in the International Journal of Community Music, General Music Today, Journal of Popular Music Education, Bulletin of the Council for Research in Music Education, and Music Education Research.

Bryan Powell is an Assistant Professor of Music Education/Music Technology at Montclair State University. Bryan is the founding co-editor of Journal of Popular Music Education, author of Popular Music Pedagogies (Routledge), and Executive Director of the Association for Popular Music Education. Bryan is the immediate past Chair for the NAfME Popular Music Education SRIG.

Gareth D Smith is Assistant Professor of Music, Music Education at Boston University. Gareth plays drums with Stephen Wheel, Dirty Blond and Black Light Bastards. His research interests include drumming, popular music education and sociology of music education. His latest book is A Philosophy of Playing Drum Kit: Magical Nexus.

John C Coggiola is a Dual Associate Professor and Chair of Music Education in the College of Visual and Performing Arts and the School of Education at Syracuse University. Dr. Coggiola is also the Director of Jazz Studies in the Setnor School of Music. Coggiola’s research focuses on aesthetic response to jazz music and is published in The Journal of Research in Music Education, The Bulletin of The Council for Research in Music Education, Contributions to Music Education, The International Association of Jazz Educators Jazz Research Proceedings Yearbook, and The Instrumentalist.

Matthew Kelsey received his Ph. D. in experimental high energy particle physics at the Large Hadron Collider in 2018 from Syracuse University. He held two postdoctoral fellowships at both Lawrence Berkeley National Laboratory and Wayne State University, where his research focused on experimental high energy particle and nuclear physics at the Relativistic Heavy-Ion Collider until 2022. Now, Matthew is a data scientist at Ford Motor Company, where he specializes in machine learning/AI techniques in the field of vehicle prognostics.