Real-Time Pattern Recognition of Symbolic Monophonic Music

This paper investigates the real-time detection of predefined monophonic patterns from the MIDI output of a digital musical instrument. This enables the development of instruments and systems for live music, which can recognize when a musician plays a certain phrase and repurpose such information to trigger external peripherals connected to the instrument. Specifically, we compare the recognition performance of Dynamic Time Warping and Recurrent Neural Network-based approaches. We employ different representation formats of musical data to optimize the efficiency of each computational method. To evaluate the algorithms, a novel dataset is introduced which includes recordings from 20 keyboard players and 20 guitar players. The evaluation focuses on the algorithms' ability to recognize patterns amid variations that impede a straightforward one-to-one comparison. The results reveal that both methods perform well in detecting up to 3 distinct patterns. However, as the number of different patterns increases up to 10, dynamic time warping exhibits a negative correlation with the recognition performance, while the recurrent neural network maintains high detection accuracy of approximately 98%. Taken together, our findings demonstrate the potential of machine learning in handling complex musical patterns in real-time, paving the way for novel applications involving smart musical instruments.