Authors:

Franklin Shang

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Automatic Music Transcription (AMT) systems face significant challenges when processing complex musical pieces, but the specific musical characteristics that contribute to transcription difficulty remain poorly understood. This research investigates the relationship between human perception of musical complexity and AMT transcription accuracy by correlating established complexity metrics with transcription performance across multiple AMT models. We analyze three key dimensions of musical complexity: polyphonic (simultaneous notes), rhythmic (timing patterns), and harmonic (chord structures). For each dimension, we implement computational metrics validated in human perception studies, including maximum polyphony, inter-onset interval variance, and chord complexity measures. These metrics are applied to datasets, and their values are correlated with transcription accuracy scores from seven different AMT models (CREPE, Sound2MIDI, MT3, Transkun, Bytedance, Basic Pitch, and ReconVAT). Our preliminary results show strong correlations between certain complexity metrics and transcription accuracy. This research contributes to understanding what makes music challenging for computational transcription systems and provides insights for improving AMT model performance. The findings can guide model selection for different types of music and inform the development of more robust transcription algorithms. Keywords: Automatic Music Transcription; Music Information Retrieval; Machine Learning

Source:

Purdue University / 2025

Topics:

No topics listed

Co-authors:

Franklin Shang