Yaafe core features

Yaafe core audio features.

Available features

AmplitudeModulation

class yaafefeatures.AmplitudeModulation

Tremelo and Grain description, according to [SE2005] and [AE2001].

AmplitudeModulation uses Envelope to describe tremolo and grain. Analyzed frequency ranges are :
  • Tremolo : 4 - 8 Hz
  • Grain : 10 - 40 Hz
For each of these ranges, it computes :
  1. Frequency of maximum energy in range
  2. Difference of the energy of this frequency and the mean energy over all frequencies
  3. Difference of the energy of this frequency and the mean energy in range
  4. Product of the two first values.
[AE2001]A.Eronen, Automatic musical instrument recognition. Master’s Thesis, Tempere University of Technology, 2001.
Parameters:
  • EnDecim (default=200): Decimation factor to compute envelope
  • blockSize (default=32768): output frames size
  • stepSize (default=16384): step between consecutive frames

Declaration example:

AmplitudeModulation EnDecim=200  blockSize=32768  stepSize=16384

See also

Envelope

AutoCorrelation

class yaafefeatures.AutoCorrelation

Compute autocorrelation coefficients ac on each frames.

ac(k) = \sum_{i=0}^{N-k-1}x(i)x(i+k)

Parameters:
  • ACNbCoeffs (default=49): Number of autocorrelation coefficients to keep
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

AutoCorrelation ACNbCoeffs=49  blockSize=1024  stepSize=512

See also

Frames

ComplexDomainOnsetDetection

class yaafefeatures.ComplexDomainOnsetDetection

Compute onset detection using a complex domain spectral flux method [CD2003].

[CD2003]C.Duxbury et al., Complex domain onset detection for musical signals, Proc. of the 6th Int. Conference on Digital Audio Effects (DAFx-03), London, UK, September 8-11, 2003
Parameters:
  • FFTLength (default=0): Frame’s length on which perform FFT. Original frame is padded with zeros or truncated to reach this size. If 0 then use original frame length.
  • FFTWindow (default=Hanning): Weighting window to apply before fft. Hanning|Hamming|None
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

ComplexDomainOnsetDetection FFTLength=0  FFTWindow=Hanning  blockSize=1024  stepSize=512

See also

Frames

Energy

class yaafefeatures.Energy

Compute energy as root mean square of an audio Frame.

en = \sqrt\frac{\sum_{i=0}^{N-1}x(i)^2}{N}

Parameters:
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

Energy blockSize=1024  stepSize=512

See also

Frames

Envelope

class yaafefeatures.Envelope

Extract amplitude envelope using hilbert transform, low-pass filtering and decimation.

Parameters:
  • EnDecim (default=200): Decimation factor to compute envelope
  • blockSize (default=32768): output frames size
  • stepSize (default=16384): step between consecutive frames

Declaration example:

Envelope EnDecim=200  blockSize=32768  stepSize=16384

See also

Frames

EnvelopeShapeStatistics

class yaafefeatures.EnvelopeShapeStatistics

Centroid, spread, skewness and kurtosis of each frame’s amplitude envelope. For more details about moments, see Shape Statistics.

Parameters:
  • EnDecim (default=200): Decimation factor to compute envelope
  • blockSize (default=32768): output frames size
  • stepSize (default=16384): step between consecutive frames

Declaration example:

EnvelopeShapeStatistics EnDecim=200  blockSize=32768  stepSize=16384

See also

Envelope

Frames

class yaafefeatures.Frames

Segment input signal into frames.

First frame has zeros on left half so that it is centered on time 0s, then consecutive frames are equally spaced. Consequently, frame i (starting from 0) is centered on sample i * stepSize.

Parameters:
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

Frames blockSize=1024  stepSize=512

LPC

class yaafefeatures.LPC

Compute the Linear Predictor Coefficients (LPC) of a signal frame. It uses autocorrelation and Levinson-Durbin algorithm. see [JM1975].

[JM1975]Makoul J., Linear Prediction: A tutorial Review, Proc. IEEE, Vol. 63, pp. 561-580, 1975.
Parameters:
  • LPCNbCoeffs (default=2): Number of Linear Predictor Coefficients to compute
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

LPC LPCNbCoeffs=2  blockSize=1024  stepSize=512

See also

AutoCorrelation

LSF

class yaafefeatures.LSF

Compute the Line Spectral Frequency (LSF) coefficients of a signal frame. Algorithm was adapted from ([TB2006], [SH1976]).

[TB2006]Tom Backstrom, Carlo Magi, Properties of line spectrum pair polynomials–A review, Signal Processing, Volume 86, Issue 11, Special Section: Distributed Source Coding, November 2006, Pages 3286-3298, ISSN 0165-1684, DOI: 10.1016/j.sigpro.2006.01.010.
[SH1976]Schussler, H., A stability theorem for discrete systems, Acoustics, Speech and Signal Processing, IEEE Transactions on , vol.24, no.1, pp. 87-89, Feb 1976
Parameters:
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

LSF blockSize=1024  stepSize=512

See also

LPC

Loudness

class yaafefeatures.Loudness

The loudness coefficients are the energy in each Bark band, normalized by the overall sum. see [GP2004] and [MG1997] for more details.

[MG1997]Moore, Glasberg, et al., A Model for the Prediction of Thresholds Loudness and Partial Loudness., J. Audio Eng. Soc. 45: 224-240, 1997.
Parameters:
  • FFTLength (default=0): Frame’s length on which perform FFT. Original frame is padded with zeros or truncated to reach this size. If 0 then use original frame length.
  • FFTWindow (default=Hanning): Weighting window to apply before fft. Hanning|Hamming|None
  • LMode (default=Relative): “Specific” computes loudness without normalization, “Relative” normalize each band so that they sum to 1, “Total” just returns the sum of Loudness in all bands.
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

Loudness FFTLength=0  FFTWindow=Hanning  LMode=Relative  blockSize=1024  stepSize=512

See also

MagnitudeSpectrum

MFCC

class yaafefeatures.MFCC

Compute the Mel-frequencies cepstrum coefficients [DM1980].

Mel filter bank is built as 40 log-spaced filters according to the following mel-scale:

melfreq = 1127 * log(1 + \frac{freq}{700})

Each filter is a triangular filter with height 2/(f_{max}-f_{min}). Then MFCCs are computed as following, using DCT II:

mfcc = dct(log(abs(fft(hanning(N).x)).MelFilterBank))

[DM1980](1, 2) S.B. Davis and P.Mermelstrin, Comparison of parametric representations for monosyllabic word recognition in continuously spoken sentences. IEEE Transactions on Acoustics, Speech and Signal Processing, 28 :357-366, 1980.
Parameters:
  • CepsIgnoreFirstCoeff (default=1): 0 keeps the first cepstral coeffcient, 1 ignore it
  • CepsNbCoeffs (default=13): Number of cepstral coefficient to keep.
  • FFTWindow (default=Hanning): Weighting window to apply before fft. Hanning|Hamming|None
  • MelMaxFreq (default=6854.0): Maximum frequency of the mel filter bank
  • MelMinFreq (default=130.0): Minimum frequency of the mel filter bank
  • MelNbFilters (default=40): Number of mel filters
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

MFCC CepsIgnoreFirstCoeff=1  CepsNbCoeffs=13  FFTWindow=Hanning  MelMaxFreq=6854.0  MelMinFreq=130.0  MelNbFilters=40  blockSize=1024  stepSize=512

See also

MagnitudeSpectrum

MagnitudeSpectrum

class yaafefeatures.MagnitudeSpectrum

Compute frame’s magnitude spectrum, using an analysis window (Hanning or Hamming), or not.

y = abs(fft(w*x))

Parameters:
  • FFTLength (default=0): Frame’s length on which perform FFT. Original frame is padded with zeros or truncated to reach this size. If 0 then use original frame length.
  • FFTWindow (default=Hanning): Weighting window to apply before fft. Hanning|Hamming|None
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

MagnitudeSpectrum FFTLength=0  FFTWindow=Hanning  blockSize=1024  stepSize=512

See also

Frames

MelSpectrum

class yaafefeatures.MelSpectrum

Compute the Mel-frequencies spectrum [DM1980].

Mel filter bank is built as 40 log-spaced filters according to the following mel-scale:

melfreq = 1127 * log(1 + \frac{freq}{700})

Each filter is a triangular filter with height 2/(f_{max}-f_{min}).

Parameters:
  • FFTWindow (default=Hanning): Weighting window to apply before fft. Hanning|Hamming|None
  • MelMaxFreq (default=6854.0): Maximum frequency of the mel filter bank
  • MelMinFreq (default=130.0): Minimum frequency of the mel filter bank
  • MelNbFilters (default=40): Number of mel filters
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

MelSpectrum FFTWindow=Hanning  MelMaxFreq=6854.0  MelMinFreq=130.0  MelNbFilters=40  blockSize=1024  stepSize=512

See also

MagnitudeSpectrum

OBSI

class yaafefeatures.OBSI

Compute Octave band signal intensity using a trigular octave filter bank ([SE2005]).

[SE2005](1, 2) S.Essid, Classification automatique des signaux audio-frequences: reconnaissance des instruments de musique. PhD, UPMC, 2005.
Parameters:
  • FFTLength (default=0): Frame’s length on which perform FFT. Original frame is padded with zeros or truncated to reach this size. If 0 then use original frame length.
  • FFTWindow (default=Hanning): Weighting window to apply before fft. Hanning|Hamming|None
  • OBSIMinFreq (default=27.5): Minimum frequency for OBSI filter.
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

OBSI FFTLength=0  FFTWindow=Hanning  OBSIMinFreq=27.5  blockSize=1024  stepSize=512

See also

MagnitudeSpectrum

OBSIR

class yaafefeatures.OBSIR

Compute log of OBSI ratio between consecutive octave.

Parameters:
  • DiffNbCoeffs (default=0): Maximum number of coeffs to keep. 0 keeps N-1 value (with N the input feature size)
  • FFTLength (default=0): Frame’s length on which perform FFT. Original frame is padded with zeros or truncated to reach this size. If 0 then use original frame length.
  • FFTWindow (default=Hanning): Weighting window to apply before fft. Hanning|Hamming|None
  • OBSIMinFreq (default=27.5): Minimum frequency for OBSI filter.
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

OBSIR DiffNbCoeffs=0  FFTLength=0  FFTWindow=Hanning  OBSIMinFreq=27.5  blockSize=1024  stepSize=512

See also

OBSI

PerceptualSharpness

class yaafefeatures.PerceptualSharpness

Compute the sharpness of Loudness coefficients, according to [GP2004].

Parameters:
  • FFTLength (default=0): Frame’s length on which perform FFT. Original frame is padded with zeros or truncated to reach this size. If 0 then use original frame length.
  • FFTWindow (default=Hanning): Weighting window to apply before fft. Hanning|Hamming|None
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

PerceptualSharpness FFTLength=0  FFTWindow=Hanning  blockSize=1024  stepSize=512

See also

Loudness

PerceptualSpread

class yaafefeatures.PerceptualSpread

Compute the spread of Loudness coefficients, according to [GP2004].

Parameters:
  • FFTLength (default=0): Frame’s length on which perform FFT. Original frame is padded with zeros or truncated to reach this size. If 0 then use original frame length.
  • FFTWindow (default=Hanning): Weighting window to apply before fft. Hanning|Hamming|None
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

PerceptualSpread FFTLength=0  FFTWindow=Hanning  blockSize=1024  stepSize=512

See also

Loudness

SpectralCrestFactorPerBand

class yaafefeatures.SpectralCrestFactorPerBand

Compute spectral crest factor per log-spaced band of 1/4 octave.

Parameters:
  • FFTLength (default=0): Frame’s length on which perform FFT. Original frame is padded with zeros or truncated to reach this size. If 0 then use original frame length.
  • FFTWindow (default=Hanning): Weighting window to apply before fft. Hanning|Hamming|None
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

SpectralCrestFactorPerBand FFTLength=0  FFTWindow=Hanning  blockSize=1024  stepSize=512

See also

MagnitudeSpectrum

SpectralDecrease

class yaafefeatures.SpectralDecrease

Compute spectral decrease accoding to [GP2004].

S_{decrease} = \frac{1}{\sum_{k=2}^{K}a_{k}}\sum_{k=2}^{K}\frac{a_{k}-a_{1}}{k-1}

Parameters:
  • FFTLength (default=0): Frame’s length on which perform FFT. Original frame is padded with zeros or truncated to reach this size. If 0 then use original frame length.
  • FFTWindow (default=Hanning): Weighting window to apply before fft. Hanning|Hamming|None
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

SpectralDecrease FFTLength=0  FFTWindow=Hanning  blockSize=1024  stepSize=512

See also

MagnitudeSpectrum

SpectralFlatness

class yaafefeatures.SpectralFlatness

Compute global spectral flatness using the ratio between geometric and arithmetic mean.

S_{flatness} = \frac{exp(\frac{1}{N}\sum_{k}log(a_{k}))}{\frac{1}{N}\sum_{k}a_{k}}

Parameters:
  • FFTLength (default=0): Frame’s length on which perform FFT. Original frame is padded with zeros or truncated to reach this size. If 0 then use original frame length.
  • FFTWindow (default=Hanning): Weighting window to apply before fft. Hanning|Hamming|None
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

SpectralFlatness FFTLength=0  FFTWindow=Hanning  blockSize=1024  stepSize=512

See also

MagnitudeSpectrum

SpectralFlatnessPerBand

class yaafefeatures.SpectralFlatnessPerBand

Compute spectral flatness per log-spaced band of 1/4 octave, as proposed in MPEG7 standard.

Parameters:
  • FFTLength (default=0): Frame’s length on which perform FFT. Original frame is padded with zeros or truncated to reach this size. If 0 then use original frame length.
  • FFTWindow (default=Hanning): Weighting window to apply before fft. Hanning|Hamming|None
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

SpectralFlatnessPerBand FFTLength=0  FFTWindow=Hanning  blockSize=1024  stepSize=512

See also

MagnitudeSpectrum

SpectralFlux

class yaafefeatures.SpectralFlux

Compute flux of spectrum between consecutives frames.

S_{flux} = \frac{\sum_{k}(a_{k}(t) - a_{k}(t-1))^2}{\sqrt{\sum_{k}a_{k}(t-1)^2}\sqrt{\sum_{k}a_{k}(t)^2}}

Parameters:
  • FFTLength (default=0): Frame’s length on which perform FFT. Original frame is padded with zeros or truncated to reach this size. If 0 then use original frame length.
  • FFTWindow (default=Hanning): Weighting window to apply before fft. Hanning|Hamming|None
  • FluxSupport (default=All): support of flux computation. if ‘All’ then use all bins (default), if ‘Increase’ then use only bins which are increasing
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

SpectralFlux FFTLength=0  FFTWindow=Hanning  FluxSupport=All  blockSize=1024  stepSize=512

See also

MagnitudeSpectrum

SpectralRolloff

class yaafefeatures.SpectralRolloff

Spectral roll-off is the frequency so that 99% of the energy is contained below. see [SS1997].

[SS1997](1, 2) E.Scheirer, M.Slaney. Construction and evaluation of a robust multifeature speech/music discriminator. IEEE Internation Conference on Acoustics, Speech and Signal Processing, p.1331-1334, 1997.
Parameters:
  • FFTLength (default=0): Frame’s length on which perform FFT. Original frame is padded with zeros or truncated to reach this size. If 0 then use original frame length.
  • FFTWindow (default=Hanning): Weighting window to apply before fft. Hanning|Hamming|None
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

SpectralRolloff FFTLength=0  FFTWindow=Hanning  blockSize=1024  stepSize=512

See also

MagnitudeSpectrum

SpectralShapeStatistics

class yaafefeatures.SpectralShapeStatistics

Compute shape statistics of MagnitudeSpectrum, (see [GR2004]).

Shape Statistics are centroid, spread, skewness and kurtosis, defined as follow:

\mu_{i} &= \frac{\sum_{n=1}^{N}f_{k}^{i}*a_{k}}{\sum_{n=1}^{N}a_{k}}\\ centroid &= \mu_{1}\\ spread &= \sqrt{\mu_{2}-\mu_{1}^{2}} \\ skewness &= \frac{2\mu_{1}^{3} - 3\mu_{1}\mu_{2} + \mu_{3}}{spread^{3}} \\ kurtosis &= \frac{-3\mu_{1}^{4} + 6\mu_{1}\mu_{2} - 4\mu_{1}\mu_{3} + \mu_{4}}{spread^{4}} - 3

[GR2004]O.Gillet, G.Richard, Automatic transcription of drum loops. in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Montreal, Canada, 2004.
Parameters:
  • FFTLength (default=0): Frame’s length on which perform FFT. Original frame is padded with zeros or truncated to reach this size. If 0 then use original frame length.
  • FFTWindow (default=Hanning): Weighting window to apply before fft. Hanning|Hamming|None
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

SpectralShapeStatistics FFTLength=0  FFTWindow=Hanning  blockSize=1024  stepSize=512

See also

MagnitudeSpectrum

SpectralSlope

class yaafefeatures.SpectralSlope

SpectralSlope is computed by linear regression of the spectral amplitude. (see [GP2004])

S_{slope} = \frac{ K\sum_{k}f_{k}a_{k} - \sum_{k}f_{k}\sum_{k}a_{k}}{K\sum_{k}f_{k}^2-(\sum_{k}a_{k})^2}

Parameters:
  • FFTLength (default=0): Frame’s length on which perform FFT. Original frame is padded with zeros or truncated to reach this size. If 0 then use original frame length.
  • FFTWindow (default=Hanning): Weighting window to apply before fft. Hanning|Hamming|None
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

SpectralSlope FFTLength=0  FFTWindow=Hanning  blockSize=1024  stepSize=512

See also

MagnitudeSpectrum

SpectralVariation

class yaafefeatures.SpectralVariation

SpectralVariation is the normalized correlation of spectrum between consecutive frames. (see [GP2004])

S_{var} = 1 - \frac{\sum_{k}a_{k}(t-1)a_{k}(t)}{\sqrt{\sum_{k}a_{k}(t-1)^2}\sqrt{\sum_{k}a_{k}(t)^2}}

[GP2004](1, 2, 3, 4, 5, 6) Geoffroy Peeters, A large set of audio features for sound description (similarity and classification) in the CUIDADO project, 2004.
Parameters:
  • FFTLength (default=0): Frame’s length on which perform FFT. Original frame is padded with zeros or truncated to reach this size. If 0 then use original frame length.
  • FFTWindow (default=Hanning): Weighting window to apply before fft. Hanning|Hamming|None
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

SpectralVariation FFTLength=0  FFTWindow=Hanning  blockSize=1024  stepSize=512

See also

MagnitudeSpectrum

TemporalShapeStatistics

class yaafefeatures.TemporalShapeStatistics

Compute shape statistics of signal frames.

Parameters:
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

TemporalShapeStatistics blockSize=1024  stepSize=512

See also

Frames

ZCR

class yaafefeatures.ZCR

Compute zero-crossing rate in frames. see [SS1997].

Parameters:
  • blockSize (default=1024): output frames size
  • stepSize (default=512): step between consecutive frames

Declaration example:

ZCR blockSize=1024  stepSize=512

See also

Frames

Available feature transforms

AutoCorrelationPeaksIntegrator

class yaafefeatures.AutoCorrelationPeaksIntegrator

Feature transform that compute peaks of the autocorrelation function, outputs peaks and amplitude.

Parameters:
  • ACPInterPeakMinDist (default=5): Minimal distance between consecutive autocorrelation peaks, expressed in lags.
  • ACPNbPeaks (default=3): Number of autocorrelation peaks to keep
  • ACPNorm (default=No): can be No|BPM|Hz. Normalize output to be expressed respectively in lag, BPM, Hz
  • NbFrames (default=60): Number of frames to integrate together
  • StepNbFrames (default=30): Number of frames to skip between two integration

Declaration example:

AutoCorrelationPeaksIntegrator ACPInterPeakMinDist=5  ACPNbPeaks=3  ACPNorm=No  NbFrames=60  StepNbFrames=30

Cepstrum

class yaafefeatures.Cepstrum

Feature transform that compute cepstrum coefficients of input feature frames. (use DCT II)

cep = dct(log(x))

Parameters:
  • CepsIgnoreFirstCoeff (default=1): 0 keeps the first cepstral coeffcient, 1 ignore it
  • CepsNbCoeffs (default=13): Number of cepstral coefficient to keep.

Declaration example:

Cepstrum CepsIgnoreFirstCoeff=1  CepsNbCoeffs=13

Derivate

class yaafefeatures.Derivate

Compute temporal derivative of input feature. The derivative is approximated by an orthogonal polynomial fit over a finite length window. (see [RR1993] p.117).

\frac{\partial x(t)}{\partial t} = \mu \sum_{k=-N}^{N}k.x(t+k) where \: \mu = \sum_{k=-N}^{N}k^2

[RR1993]L.R.Rabiner, Fundamentals of Speech Processing. Prentice Hall Signal Processing Series. PTR Prentice-Hall, 1993.
Parameters:
  • DO1Len (default=4): Horizon used to compute order 1 derivative.
  • DO2Len (default=1): Horizon used to compute order 2 derivative. Useless if DOrder=1.
  • DOrder (default=1): Order of the derivative to compute.

Declaration example:

Derivate DO1Len=4  DO2Len=1  DOrder=1

HistogramIntegrator

class yaafefeatures.HistogramIntegrator

Feature transform that compute histogram of input values

Parameters:
  • HInf (default=0): Minimal value to take into consideration
  • HNbBins (default=10): Nb bins of histogram
  • HSup (default=1): Maximal value to take into consideration
  • HWeighted (default=0): Set it to 1 if input values are weighted. If 1, input is considered to be a list of couple (value,weight).
  • NbFrames (default=60): Number of frames to integrate together
  • StepNbFrames (default=30): Number of frames to skip between two integration

Declaration example:

HistogramIntegrator HInf=0  HNbBins=10  HSup=1  HWeighted=0  NbFrames=60  StepNbFrames=30

SlopeIntegrator

class yaafefeatures.SlopeIntegrator

Feature transform that compute the slope of input feature over the given number of frames.

Parameters:
  • NbFrames (default=60): Number of frames to integrate together
  • StepNbFrames (default=30): Number of frames to skip between two integration

Declaration example:

SlopeIntegrator NbFrames=60  StepNbFrames=30

StatisticalIntegrator

class yaafefeatures.StatisticalIntegrator

Feature transform that compute the temporal mean and variance of input feature over the given number of frames.

Parameters:
  • NbFrames (default=60): Number of frames to integrate together
  • SICompute (default=MeanStddev): if ‘MeanStddev’ then compute mean and standard deviation, if ‘Mean’ compute only mean, if ‘Stddev’ compute only stantard deviation.
  • StepNbFrames (default=30): Number of frames to skip between two integration

Declaration example:

StatisticalIntegrator NbFrames=60  SICompute=MeanStddev  StepNbFrames=30

Table Of Contents

Navigation

© Copyright 2009-2011 Institut Télécom - Télécom Paristech. Created using Sphinx 1.0.7.
Project hosted on Get yaafe at SourceForge.net. Fast, secure and Free Open Source software downloads