In this study we discuss some of the limitations of Gaussian humanization and consider ways in which the articulation patterns exhibited by percussionists can be emulated using a probabilistic model. Prior and likelihood functions are derived from a dataset of professional drummers to create a series of empirical distributions. These are then used to independently modulate the onset locations and amplitudes of a quantized sequence, using a recursive Bayesian framework. Finally, we evaluate the performance of the model against sequences created with a Gaussian humanizer and sequences created with a Hidden Markov Model (HMM) using paired listening tests. We are able to demonstrate that probabilistic models perform better than instantaneous Gaussian models, when evaluated using a 4/4 rock beat at 120 bpm.
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.