MalletMetrics
Automatic performance assessment for marimba
A recording-based system that transcribes marimba performances, aligns them with a reference musical sequence, and evaluates pitch, timing, and dynamics to produce musician-facing feedback.
The question
Useful performance feedback needs to account for more than whether the right notes were played. MalletMetrics explores computational assessment of timing, pitch, and dynamics, with the goal of giving percussionists feedback they can act on.
From recording to feedback
The transcription stage uses YourMT3+. TheGlueNote then aligns the transcribed performance with a reference sequence. Median-offset normalization accounts for global timing offsets, while outlier filtering reduces anomalous alignment errors.
For dynamics, RMS amplitude and spectral centroid feed a K-nearest-neighbors regression model. Calibration recordings cover labeled playing levels from approximately piano through fortissimo. GPT-4o mini converts the computational evaluations into pedagogical feedback.
Evaluation approach
The study included 50 recorded performances from 10 performers ranging from novice to professional, alongside 15 expert evaluators. The evaluation examined alignment between system feedback and expert musical assessment.
Quantitative results are omitted here until they can be checked against the manuscript. Study scale alone does not establish assessment quality.
Limits & next questions
The system processes complete recordings afterward; it is not a real-time performance tool. Transcription and alignment errors can affect downstream assessments, and the relationship between acoustic features and perceived dynamics warrants careful calibration.
The broader research question is how to make computational assessments musically useful while preserving the judgment and agency of performers and teachers.