Drum Sound Recognition Based On Rhythm

Timbre information and Rhythm information are both important for drum recognition. Time domain and time-frequency domain template adaption methods, feature classification method of drum sound detection only consider about drum timbre, so the accuracy and anti-interference are yet to be improved. And due to the separation quality and the accuracy of sound source number estimation, the results of source separation method are not pleasant. This paper combines onset detection and rhythm detection with drum sound recognition; respectively work on drum sound recognition based on rhythm and drum sound recognition based on rhythm and/or timbre. The effectiveness of the two methods is verified by experiment.This paper also describes the approach to detect drum sounds from polyphonic music based on drum energy and drum loop. Considering that the energy of drum is much higher than that of harmonic instruments, the possible drum onsets are indicated by efficiently combining accent signal, NAE beat spectrum, and accent spectrum into a clearer curve for recognition. The drum segments are then recognized using the rhythm information. The experiments show that the proposed algorithm improves the accuracy and the recall rate compared to the template iteration methods of both time-frequency domain and time domain, the feature classification method, or the source separation method.Finally, this paper proposes a recognition prototype using both timbre and rhythm, which is compared to all the above mentioned methods. Experiment results show that though less robust than the method using drum energy, this prototype is far better than the methods using template iteration in either time domain or time-frequency domain and the method using feature classification.