Perceptual tempo estimation using GMM-regression

Abstract

Most current tempo estimation algorithms suffer from the so-called octave estimation problems (estimating twice, thrice, half or one-third of a reference tempo). However, it is difficult to qualify an error as octave error without a clear definition of what is the reference tempo. For this reason, and given that tempo is mostly a perceptual notion, we study here the estimation of perceptual tempo. We consider the perceptual tempo as defined by the results of the largescale experiment made at Last-FM in 2011. We assume that the perception of tempo is related to the rate of variation of four musical attributes: the variation of energy, of harmonic changes, of spectral balance and short-term-eventrepetitions. We then propose the use of GMM-Regression to find the relationship between the perceptual tempo and the four musical attributes. In an experiment, we show that the estimation of the tempo provided by GMM-Regression over these attributes outperforms the one provided by a stateof-the-art tempo estimation algorithm. For this task GMMRegression also largely outperforms SVM-Regression. We finally study the estimation of three perceptual tempo classes ("Slow", "In Between", "Fast") using both GMM-Regression and SVM-Classification.

Predicting Agreement and Disagreement in the Perception of Tempo

“…We successfully validated these four functions in [15] for the estimation of perceptual tempo (in the case f (a, ∀u) =t). We briefly summarized these functions in part 2.1.…”

“…As opposed to previous studies, the work of Peeters et al [15] is one of the few to study perceptual tempo estimation on real annotated perceptual tempo data (derived from the perceptual experiment performed at Last-FM in 2011). They propose four feature sets to describe the audio content and propose the use of GMM-Regression [3] to model the relationship between the audio features and the perceptual tempo.…”

“…Rather than estimating the chord succession, we estimate the rate at which segments of stable harmonic content vary. In [15] we proposed to represent this using Chroma variations over time. The variation is computed by convolving a Chroma SelfSimilarity-Matrix with a novelty kernel [6] whose length represent the assumption of chord duration.…”

“…Each of these tempo estimation is made using our previous GMMRegression methods as described in [15]. Each track a is then represented by a 4-dim feature vector where each dimension represent the prediction of tempo using a specific feature set: [t ener ,t sim ,t specbal ,t harmo ].…”

An effective, simple tempo estimation method based on self-similarity and regularity

“…The "feature" was the lowest tempo candidate of the pair and the desired outcome was whether to double or not. This can be viewed as an extremely simplified version of the more complex machine learning approaches that have been proposed in the literature [16,22,15]. Bold numbers indicate the best-performing algorithm for this dataset.…”