SDTDemix.h: Transient/tonal/residual components separator

Typedefs
typedef struct SDTDemix	SDTDemix
	Opaque data structure for the percussive/harmonic/residual components separator.
typedef struct SDTDemix	SDTDemix
	Opaque data structure for the percussive/harmonic/residual components separator.

Functions
SDTDemix *	SDTDemix_new (int size, int radius)
	Object constructor. More...
void	SDTDemix_free (SDTDemix *x)
	Object destructor. More...
void	SDTDemix_setOverlap (SDTDemix *x, double f)
	Sets the window overlapping factor. More...
void	SDTDemix_setNoiseThreshold (SDTDemix *x, double f)
	Sets the noise threshold. More...
void	SDTDemix_setTonalThreshold (SDTDemix *x, double f)
	Sets the tornal threshold. More...
void	SDTDemix_dsp (SDTDemix *x, double *outs, double in)
	Signal processing routine. Call this function at sample rate to separate an arbitrary signal into its percussive/harmonic/residual components. More...

Detailed Description

This algorithm looks for vertical and horizontal structures in the spectrogram to separate an arbitrary audio signal into its percussive (transients), harmonic (sustained tones) and residual (noise) components. It is based on a technique called structure tensor, frequently used in image processing for edge and corner detection or to estimate the orientation of an object.

The structure tensor can be thought as a smoothed gradient of the spectrogram, which describes the consistency and direction of changes in the energy content of each bin. The anisotropy (consistency) and direction descriptors extracted from the structure tensor are used to classify the spectrogram bins into three categories: Bins which do not exhibit a particular gradient direction (low anisotropy) become part of the residual, noisy component; Bins which tend to have a vertical orientation (high anisotropy, high direction) are included in the percussive component; Bins with a mostly horizontal orientation (high anisotropy, low direction) fall into the harmonic component.

This percussive/harmonic/residual separation is suitable to separate attacks, decays and noise from a musical signal, or to isolate myoelastic (percussive), phonatory (harmonic) and turbulent (noisy) activities from a vocal signal. In particular, the algorithm can be used as a preprocessing step to improve the results of the myoelastic detector, pitch tracker and spectral moments extractor present in the analysis part of the Sound Design Toolkit.

Function Documentation

void SDTDemix_dsp	(	SDTDemix *	x,
		double *	outs,
		double	in
	)

Signal processing routine. Call this function at sample rate to separate an arbitrary signal into its percussive/harmonic/residual components.

Parameters

[out]	outs	Pointer to a 3-elements output array. outs[0] is the percussive component, outs[1] is the harmonic component and outs[2] is the residual.
[in]	in	Input sample

void SDTDemix_free

(

SDTDemix *

x

)

Object destructor.

Parameters

[in]

x

Pointer to the instance to destroy

SDTDemix * SDTDemix_new	(	int	size,
		int	radius
	)

Object constructor.

Parameters

[in]	size	Analysis window length, in samples
[in]	radius	Smoothing kernel radius, in samples

Returns

Pointer to the new instance

void SDTDemix_setNoiseThreshold	(	SDTDemix *	x,
		double	f
	)

Sets the noise threshold.

Parameters

[in]

f

Amount of signal falling into the residual category

void SDTDemix_setOverlap	(	SDTDemix *	x,
		double	f
	)

Sets the window overlapping factor.

Parameters

[in]

f

Window overlapping factor

void SDTDemix_setTonalThreshold	(	SDTDemix *	x,
		double	f
	)

Sets the tornal threshold.

Parameters

[in]

f

Amount of non-residual falling into the tonal category