Abstract
Traditional drum transcription systems often rely solely on short-term acoustic information, neglecting the inherent structural and rhythmic syntax of music. This paper proposes two post-processing strategies to incorporate structural "sequence models": a supervised generalized N-gram model and an unsupervised complexity-minimization algorithm. By ensuring transcribed drum patterns are musically coherent and repetitive, these methods significantly improve transcription accuracy over standard acoustic baselines.
Introduction
Most drum transcription systems focus on identifying a drum hit based solely on the sound at a specific moment. However, music is not random; it follows a syntax. Drummers play patterns that are repetitive, style-specific, and structurally organized—for instance, a kick drum on every beat in disco or a specific snare fill at the end of a phrase. Acoustic classifiers often make mistakes that "break" this musical structure. This research addresses the gap between raw acoustic detection and high-level structural rules, aiming to "correct" transcriptions to make them more musically logical.
Methodology
The authors propose two distinct approaches operating on a symbolic representation where drum hits are quantized to a pulse grid (tatum):
1. Supervised: Generalized N-gram Models
Standard N-grams look at the immediate past ($n-1, n-2$). The authors introduce a time support $T$, allowing the model to look back at specific musical intervals, such as one beat or one bar ago. They evaluated several training strategies: Generic (all drummers/styles), Drummer-dependent, Style-dependent (e.g., rock vs. funk), and Individual (knowing the specific pattern in advance).
2. Unsupervised: Complexity-Based Optimization
This novel approach assumes that a "correct" drum transcription should be simple to describe (highly repetitive). The system uses the SEQUITUR algorithm to estimate the Kolmogorov complexity of a sequence by representing it as a context-free grammar. Transcription is framed as an optimization problem: finding a sequence that maximizes the penalized likelihood—balancing a high acoustic score with low complexity. A Genetic Algorithm is employed to evolve the best transcription within this search space.
Experimental Setup and Results
The study utilized the ENST-drums corpus, featuring "minus-one" sequences (drum tracks with accompaniment) across various genres. Performance was measured using the F-measure for three instruments: Bass Drum (BD), Snare Drum (SD), and Hi-Hat (HH).
Key Insights
- The Syntax of Rhythm: Just as language models improve speech recognition, drum transcription benefits from a "rhythmic grammar." Generalized N-grams capture the periodic nature of drumming far better than standard Markov chains.
- Complexity as a Proxy for Correctness: The unsupervised method demonstrates that errors made by acoustic classifiers typically increase sequence complexity. By "simplifying" the sequence while respecting acoustic data, the system naturally fixes errors.
- The "Confident Error" Problem: Acoustic classifiers often assign high probability to the wrong instrument rather than just being "noisy." This makes it difficult for sequence models to overrule the acoustic data without a very high penalty term.
Conclusion and Future Work
Both supervised and unsupervised methods yielded significant improvements. The unsupervised complexity-based method, in particular, proved that structural regularity is a powerful signal even without a training corpus. However, limitations remain, including the high computational cost of Genetic Algorithms and the dependency on the initial reliability of the acoustic classifier. Future work will explore more efficient optimization heuristics and apply complexity-based regularization to broader Music Information Retrieval (MIR) tasks.
Summary updated