Smart Drums: Comparing Audio and MIDI in Embedded Real-Time Drum Pattern Recognition

This paper compares dynamic time warping (DTW) and recurrent neural networks (RNNs) for real-time drum pattern recognition on embedded hardware using both MIDI and audio representations. Motivated by Smart Drums—electronic drum kits with embedded intelligence requiring low-latency, robust pattern detection to repurpose drum phrases as high-level control signals during live performances—a dataset was recorded from 20 drummers, comprising 200 distinct drum patterns, each with ten expressive variations (2,000 performance instances), recorded in MIDI and audio. DTW serves as an alignment-based baseline using pattern-specific thresholds applied to MIDI matrices and MFCC-based audio features. The RNN employs a stacked recurrent architecture trained on synthetic pattern variations generated by a rule-based model. Experiments assess accuracy (precision, recall, and F1-score), robustness to expressive variation, and embedded performance (latency, CPU, and RAM usage) on a Raspberry Pi 4 across sets of one, three, and ten simultaneous patterns. Results show that the RNN consistently attains higher F1-scores and lower inference times than DTW in both modalities, with the performance gap widening as the number of simultaneous patterns increases. While MIDI generally yields higher detection accuracy than audio, the RNN maintains robust performance in both domains, making it better suited for deployment on Smart Drums.