Overview

Solution and Impact

We designed and developed custom training and testing software. After training, users were able to reliably identify the composer of a piano clip in the test. So we showed experimentally that composer styles exist, and that they can be learned.

Problem

Music experts talk about composers having unique styles of composition. This is a key piece of how music is evaluated, sold, and taught. So it is important that there is strong evidence for composers’ styles. But there was no empirical evidence of this in the academic literature. So we investigated whether composers have unique, learnable styles of composition.

Methods

Literature review, Experiment Design, Software Design, Survey, Statistical Data Analysis

• Part 1: Research and Methodology - Literature Review

• We reviewed and synthesized existing research to better understand the problem and evaluate initial hypotheses.

• Part 2: Research and Methodology - Experiment

  We used a quasi-experimental study design: We tested participants, trained them to distinguish composers’ styles, then tested again.

• Part 3: Design - Learning Program

  We designed a custom learning program for the training. We prioritized functionality over aesthetics because the software was for research purposes only.

• Part 4: Design - Composer Style Test

  We designed a new test of composers’ styles for before and after training.

• Part 5: Results and Data Analysis - Statistics and Data Visualizations

  We statistically analyzed our data to answer our research question.

Conclusions and Impact

Reflection

Our initial hypotheses, based on team members’ experience with studying and listening to music, were:

1. Composers have distinct styles of composition.

2. Composer style is hard to learn.

We explored the validity of these hypotheses through reviewing research on music learning, musical style, auditory perception, and perceptual learning. We read and discussed papers in our weekly meeting, identifying how each contributed to our understanding of our topic. Then we synthesized the research, identifying insights and gaps in the literature:

• Humans can learn complex patterns (e.g. Thai, Krasne, & Kellman, 2015), including musical themes (Java, Kaminska, & Gardiner, 1995)

• Musical period style has been extensively studied by psychologists (e.g. Hasenfus, Martindale, & Birnbaum, 1983)

• The only 2 studies relevant to learning composer style showed some learning (Crump, 2002; Tyler, 1946), but did not separate composer style from musical form (i.e. sonata, rondo, waltz, etc.)

We identified a gap in the research: the research question of whether composer style exists and can be learned. There were two studies that suggested composer style may exist, but both did not separate the effects of composer from the effects of musical form. There was no direct evidence about the difficulty of learning composer style, though the lack of research suggests that it may be difficult.

This gap may be hard to fill, requiring extensive training of participants and careful study design. But it is also a theoretically and practically important gap, so we decided to do our best to fill it.

(back to Approach)

Study Design and Procedure

We decided to use a quasi-experimental pretest-posttest design: test participants, train them, and then test them again (see Figure 2). The study was 9 days long, beginning and ending with a lab session, so it would have been difficult to recruit a large enough number of participants to have a control group in addition to the experimental group.

We tested participants before and after training, so that we could use the first Composer Style Test as a baseline of untrained test performance. Even though we could not compare performance to a control group, we could compare trained performance to baseline.

Participants

Recruitment: Participants were recruited through an online participant recruitment platform, which allows potential participants to sign up for studies they are interested in and available for.

Demographics: 41 UCLA undergraduate students (19 men, 24 women) volunteered.

Survey

We wrote a short open-ended survey to learn about our participants’ musical experience. We asked about their formal musical training and instruments that they played. On average, participants had about 7 years of musical instrument training, typically on the piano. 17 participants did not have music training before the study.

(back to Approach)

Content

We collected 100 15-second clips of piano music for each composer. We discussed various ways of selecting clips, and eventually decided to randomly selected a start time for each clip to minimize bias.

We wanted our users to learn the composers’ styles, and not just their favorite instrumentation, so we controlled for that variable by only using solo piano music. We selected the above composers because we found many high-quality recordings of several different solo piano pieces for each composer.

We also considered other ways composers and clips of music can be different from each other: volume, tempo, major or minor key, emotional quality (happy or sad or calm), etc. We wanted users to learn the essence of each composer’s style, so we included clips across a range of these variables for each composer.

Learning Algorithm

Users did not learn for a fixed amount of time. Instead, the learning was adapted to each user’s performance using Insight Learning Technology’s learning algorithm: Each user’s trial-by-trial performance was tracked for each composer. Users completed the Learning Program when they had answered correctly on most of the last several trials for each composer.

The learning algorithm was designed for learning visual materials, which are presented across space. Unlike visual materials, music is presented across time. We adapted Insight’s learning algorithm by removing response time from it, and by adding a delay before advancing to the next trial for incorrect answers. We wanted to make sure that participants heard most of the clip and had time to process it, so if they answered incorrectly, they had to continue listening (see Figure 5).

User Flows

The learning program had two parts: a user signup and login flow, and a flow for learning trials with feedback (see Figure 5).

Sign Up and Log In

Users set up accounts by creating a unique username and a password, and entering an email. Account creation required entering an email so that users could be contacted. Users accessed the learning program remotely by logging in with their username and password between laboratory sessions.

Learning Trials

The learning program was composed of learning trials (see video). In each trial, users heard a clip, saw “Which composer?”, clicked on a button labeled with a composer’s name, and then received feedback. The order of trials was customized for each user by the learning algorithm. We randomized the order of composer name buttons on each trial so that users learned the composer’s names and not a particular button location.

Trial Structure (see Figure 4)

1. Listen to clip

2. Attempt to identify composer of the clip by clicking on the composer’s name

3. Receive feedback on trial accuracy - feedback both helps users learn and is also motivating

   • If the answer was correct, give feedback that user was correct, then continue to the next trial - assume that the user knows the composer

   • If the answer was incorrect, show the correct answer and listen to (at least) 10 total seconds of the clip, then continue to the next trial - assume that the user does not know the composer and needs to get some more musical information (which is presented over time) from the clip

   • If the user did not respond in 30 seconds, show the correct answer and count as incorrect, then continue to the next trial - assume that the user does not know the composer even after hearing the whole clip

(back to Approach)

Content

We wanted to test users’ ability to identify the composer of a short clip of piano music. We tested whether users learned the 4 composers in the Learning Program training.

We wanted to have a high chance of showing learning, so we decided to test users on more than just the composers included in the training. We debated exactly who else to include in the test, then decided to include another composer in each trained musical period (Baroque, Romantic) and one composer each from two other periods (pre-Baroque and post-Romantic, see Figure 5).

This was our reasoning: If users learned the styles of the composers in the training but confused the composers’ names, then users would show learning by distinguishing trained composers from the other composer in each trained period. If users learned the period styles only, then users would still show learning by distinguishing trained periods from untrained periods.

We had 24 clips in the test. We included more than one clip per composer to get more data points to combine into better estimates of learning, and so that users could not just responding to one particularly distinctive clip.