Rhythm and pitch detection

The start of each "event"

May or may not fall on a beat

But not all beats have onsets necessarily

When off-beat stresses are common

Based on musical structure recognition (e.g. melodic repetition)

Primarily for symbolic representations

Finds periodicities in data

Best for tempo. Needs additional techniques to find beat phase (i.e. where the strong beats are, and the strength of current beat)

Can be done on different representations

The signal is fed through resonating filters, the highest of them determines the tempo

Histograms of inter-onset intervals

First-order (only between an onset and the next one), or higher order to compare more distant beats

"While there is a great deal of variation in the actual implementation and often the finer details are left unreported, the basic philosophy is to have a number of agents or hypotheses which track independently; these maintain an expectation of the underlying beat process and are scored with their match to the data. Low-scoring agents are killed while high-scoring ones may be branched to cover differing local hypotheses. At the end of the signal, the agent with the highest score wins and is chosen."

"Probabilistic approaches can have similarities to multiple agent architectures in that the models underlying each can be very similar. However, while the latter use a number of discrete agents which assess themselves in isolation, probabilistic models maintain distributions of all parameters and use these to arrive at the best hypothesis."

"This allows the use of standard estimation procedures such as the Kalman filter [41], Markov chain Monte Carlo (MCMC) methods [208], or sequential Monte Carlo (particle filtering) algorithms [22]"

In this version, the pitches are presented in the order of the circles of fifths