NoveltyCurve¶

streaming mode | Rhythm category

Inputs¶

frequencyBands (vector_real) - the frequency bands

Outputs¶

novelty (real) - the novelty curve as a single vector

Parameters¶

frameRate (real ∈ [1, ∞), default = 344.531) :
the sampling rate of the input audio

normalize (bool ∈ {true, false}, default = false) :
whether to normalize each band’s energy

weightCurve (vector_real, default = []) :
vector containing the weights for each frequency band. Only if weightCurveType==supplied

weightCurveType (string ∈ {flat, triangle, inverse_triangle, parabola, inverse_parabola, linear, quadratic, inverse_quadratic, supplied}, default = inverse_quadratic) :
the type of weighting to be used for the bands novelty

Description¶

This algorithm computes the “novelty curve” (Grosche & Müller, 2009) onset detection function. The algorithm expects as an input a frame-wise sequence of frequency-bands energies or spectrum magnitudes as originally proposed in [1] (see FrequencyBands and Spectrum algorithms). Novelty in each band (or frequency bin) is computed as a derivative between log-compressed energy (magnitude) values in consequent frames. The overall novelty value is then computed as a weighted sum that can be configured using ‘weightCurve’ parameter. The resulting novelty curve can be used for beat tracking and onset detection (see BpmHistogram and Onsets).

Notes:

Recommended frame/hop size for spectrum computation is 2048/1024 samples (44.1 kHz sampling rate) [2].
Log compression is applied with C=1000 as in [1].
Frequency bands energies (see FrequencyBands) as well as bin magnitudes for the whole spectrum can be used as an input. The implementation for the original algorithm [2] works with spectrum bin magnitudes for which novelty functions are computed separately and are then summarized into bands.
In the case if ‘weightCurve’ is set to ‘hybrid’ a complex combination of flat, quadratic, linear and inverse quadratic weight curves is used. It was reported to improve performance of beat tracking in some informal in-house experiments (Note: this information is probably outdated).

References:

Grosche, P. & Müller, M. (2009). A mid-level representation for capturing dominant tempo and pulse information in music recordings. International Society for Music Information Retrieval Conference (ISMIR 2009).
Tempogram Toolbox (Matlab implementation), http://resources.mpi%2Dinf.mpg.de/MIR/tempogramtoolbox

Source code¶

C++ source code

C++ header file