Skip to content

Commit

Permalink
Work on Octave frequency scale
Browse files Browse the repository at this point in the history 
Issue #332
  • Loading branch information
@towsey
towsey committed Aug 18, 2020
1 parent 5031e9a commit 5f26e3f
Show file tree
Hide file tree
Showing 3 changed files with 209 additions and 173 deletions.
110 changes: 42 additions & 68 deletions src/AudioAnalysisTools/DSP/OctaveFreqScale.cs
View file
Original file line number Diff line number Diff line change
Expand Up @@ -58,20 +58,41 @@ public static FrequencyScale GetStandardOctaveScale(FrequencyScale scale)
}

/// <summary>
/// Returns a decibel spectrogram after converting frequency scale from linear to octave.
/// This method is octave frequency scale equivalent of MFCCStuff.DecibelSpectra(dspOutput.AmplitudeSpectrogram, dspOutput.WindowPower, sampleRate, epsilon)
/// The MFCCStuff method is the proper way to convert amplitude spectrogram to decibels.
/// It converts an amplitude spectrogram to a power spectrogram having specified frequency scale.
/// It transforms the amplitude spectrogram in the following steps:
/// (1) It removes the DC row or bin 0 iff there is odd number of spectrogram bins. ASSUMPTION: Bin count should be power of 2 from FFT.
/// (2) Then reduce the linear scale to an octave scale depending on the sr and required number of bins or filters.
/// (3) It converts spectral amplitudes to power, normalising for window power and sample rate.
/// The window contributes power to the signal which must subsequently be removed from the spectral power. Calculate power per sample.
/// See notes in the MFCCStuff.DecibelSpectra for further exp[lanaitons. These normalisations were adapted from MatLab MFCC code.
/// (4) It converts the power spectrogram to decibels.
/// </summary>
/// <param name="inputSpgram">An amplitude spectrogram with linear freqeuncy scale.</param>
/// <param name="freqScale">The octave frequency scale to be used.</param>
/// <returns>decibel spectrogram having octave scale.</returns>
public static double[,] ConvertAmplitudeSpectrogramToDecibelOctaveScale(double[,] inputSpgram, FrequencyScale freqScale, double epsilon)
/// <param name="amplitudeM">The amplitude spectrogram.</param>
/// <param name="windowPower">FFT window power comes from DSP_Frames.WindowPower.</param>
/// <param name="sampleRate">of the original signal.</param>
/// <param name="epsilon">dependson bit-rate of the original signal.</param>
/// <param name="freqScale">In this case an octave frquency scale.</param>
/// <returns>The decibel spectrogram.</returns>
public static double[,] ConvertAmplitudeSpectrogramToFreqScaledDecibels(double[,] amplitudeM, double windowPower, int sampleRate, double epsilon, FrequencyScale freqScale)
{
//square the values to produce power spectrogram
var dataMatrix = MatrixTools.SquareValues(inputSpgram);
int frameCount = amplitudeM.GetLength(0);
int binCount = amplitudeM.GetLength(1);

//convert spectrogram to octave scale
var newMatrix = ConvertLinearSpectrogramToOctaveFreqScale(dataMatrix, freqScale);
newMatrix = MatrixTools.Power2DeciBels(newMatrix, epsilon, out var min, out var max);
return newMatrix;
if (binCount.IsOdd())
{
// remove the DC freq bin 0.
amplitudeM = MatrixTools.Submatrix(amplitudeM, 0, 1, frameCount - 1, binCount - 1);
}

var octaveScaleM = ConvertLinearSpectrogramToOctaveFreqScale(amplitudeM, freqScale);
var powerSpectrogram = ConvertAmplitudeToPowerSpectrogram(octaveScaleM, windowPower, sampleRate);

// Convert the power values to log using: dB = 10*log(power)
var powerEpsilon = epsilon* epsilon / windowPower / sampleRate;
var decibelSpectrogram = MatrixTools.SpectrogramPower2DeciBels(powerSpectrogram, powerEpsilon, out var min, out var max);
return decibelSpectrogram;
}

/// <summary>
Expand Down Expand Up @@ -116,73 +137,26 @@ public static FrequencyScale GetStandardOctaveScale(FrequencyScale scale)
return octaveSpectrogram;
}

public static double[,] DecibelSpectra(double[,] amplitudeM, double windowPower, int sampleRate, double epsilon, FrequencyScale freqScale)
{
double[,] powerSpectra = PowerSpectra(amplitudeM, windowPower, sampleRate, epsilon, freqScale);

// Convert the power values to log using: dB = 10*log(power)
var decibelSpectra = MatrixTools.Power2DeciBels(powerSpectra, epsilon, out var min, out var max);
return decibelSpectra;
}

public static double[,] AmplitudeSpectra(double[,] amplitudeM, double windowPower, int sampleRate, double epsilon, FrequencyScale freqScale)
{
double[,] powerSpectra = PowerSpectra(amplitudeM, windowPower, sampleRate, epsilon, freqScale);

// Convert the power values back to amplitude by taking the square root.
var amplitudeSpectra = MatrixTools.SquareRootOfValues(powerSpectra);
return amplitudeSpectra;
}

/// <summary>
/// Converts an amplitude spectrogram to a power spectrogram having an octave frequency scale.
/// It transforms the amplitude spectrogram in the following steps:
/// (1) It removes the DC row or bin 0 iff there is odd number of spectrogram bins. ASSUMPTION: Bin count should be power of 2 from FFT.
/// (1) It converts spectral amplitudes to power, normalising for window power and sample rate.
/// The window contributes power to the signal which must subsequently be removed from the spectral power. Calculate power per sample.
/// See notes in the MFCCStuff.DecibelSpectra for further exp[lanaitons. These normalisations were adapted from MatLab MFCC code.
/// (2) Then reduce the linear scale toan octave scale depending on the sr and required number of bins or filters.
/// </summary>
/// <param name="amplitudeM"> the amplitude spectra. </param>
/// <param name="windowPower">value for window power normalisation.</param>
/// <param name="sampleRate">to NormaliseMatrixValues for the sampling rate.</param>
/// <param name="epsilon">small value to avoid log of zero.</param>
/// <param name="freqScale">the kind of frequency scale.</param>
public static double[,] PowerSpectra(double[,] amplitudeM, double windowPower, int sampleRate, double epsilon, FrequencyScale freqScale)
public static double[,] ConvertAmplitudeToPowerSpectrogram(double[,] amplitudeM, double windowPower, int sampleRate)
{
int frameCount = amplitudeM.GetLength(0);
int binCount = amplitudeM.GetLength(1);

double minPow = epsilon * epsilon / windowPower / sampleRate;
double minPow2 = epsilon * epsilon * 2 / windowPower / sampleRate;
// init the octave scaled spectrogram as a matrix of spectra
double[,] powerSpectrogram = new double[frameCount, binCount];

if (binCount.IsOdd())
{
// remove the DC freq bin 0.
amplitudeM = MatrixTools.Submatrix(amplitudeM, 0, 1, frameCount - 1, binCount - 1);
}

// init the spectrogram as a matrix of spectra
double[,] powerSpectra = new double[frameCount, binCount];

// first square the values to calculate power.
// Square the values to calculate power.
// Must multiply by 2 to accomodate two spectral components, ie positive and neg freq.
for (int j = 0; j < binCount - 1; j++)
{
//foreach time step or frame
for (int i = 0; i < frameCount; i++)
{
if (amplitudeM[i, j] < epsilon)
{
powerSpectra[i, j] = minPow2;
}
else
{
powerSpectra[i, j] = amplitudeM[i, j] * amplitudeM[i, j] * 2 / windowPower / sampleRate;
}
} //end of all frames
} //end of all freq bins
powerSpectrogram[i, j] = amplitudeM[i, j] * amplitudeM[i, j] * 2 / windowPower / sampleRate;
}
}

/*
//calculate power of the Nyquist freq bin - last column of matrix
//foreach time step or frame
for (int i = 0; i < frameCount; i++)
Expand All @@ -197,9 +171,9 @@ public static FrequencyScale GetStandardOctaveScale(FrequencyScale scale)
powerSpectra[i, binCount - 1] = amplitudeM[i, binCount - 1] * amplitudeM[i, binCount - 1] / windowPower / sampleRate;
}
}
*/

powerSpectra = ConvertLinearSpectrogramToOctaveFreqScale(powerSpectra, freqScale);
return powerSpectra;
return powerSpectrogram;
}

public static int[,] GetGridLineLocations(int nyquist, int linearBound, int[,] octaveBinBounds)
Expand Down
30 changes: 8 additions & 22 deletions src/AudioAnalysisTools/Indices/IndexCalculate.cs
View file
Original file line number Diff line number Diff line change
Expand Up @@ -166,17 +166,13 @@ public static IndexCalculateResult Analysis(
{
// Only allow one octave scale at the moment, OctaveDataReduction.
// This is used to produced a reduced set of indices for content recognition.
// TODO: Much work required to generalise this for other Octave scales. Need to put scale type in the config file and and set up recovery here.
// TODO: Much work required to generalise this for other Octave scales.
// Need to put scale type in the config file and and set up recovery here.
freqScale = new FrequencyScale(FreqScaleType.OctaveDataReduction);

// Recalculate the spectrogram according to octave scale. This option works only when have high SR recordings.
dspOutput1.AmplitudeSpectrogram = OctaveFreqScale.AmplitudeSpectra(
dspOutput1.AmplitudeSpectrogram,
dspOutput1.WindowPower,
sampleRate,
epsilon,
freqScale);
dspOutput1.NyquistBin = dspOutput1.AmplitudeSpectrogram.GetLength(1) - 1; // ASSUMPTION!!! Nyquist is in top Octave bin - not necessarily true!!
// Calculate the octave scaled spectrogram.
dspOutput1.AmplitudeSpectrogram = OctaveFreqScale.ConvertLinearSpectrogramToOctaveFreqScale(dspOutput1.AmplitudeSpectrogram, freqScale);
dspOutput1.NyquistBin = dspOutput1.AmplitudeSpectrogram.GetLength(1) - 1; // ASSUME Nyquist is in top Octave bin.
}
else if (melScale)
{
Expand All @@ -190,10 +186,6 @@ public static IndexCalculateResult Analysis(
maxFreq);

dspOutput1.NyquistBin = dspOutput1.AmplitudeSpectrogram.GetLength(1) - 1;

// TODO: This doesn't make any sense, since the frequency width changes for each bin. Probably need to set this to NaN.
// TODO: Whatever uses this value below, should probably be changed to not be depending on it.
dspOutput1.FreqBinWidth = sampleRate / (double)dspOutput1.AmplitudeSpectrogram.GetLength(1) / 2;
}

// NOW EXTRACT SIGNAL FOR BACKGROUND NOISE CALCULATION
Expand All @@ -215,15 +207,9 @@ public static IndexCalculateResult Analysis(
// If necessary, recalculate the spectrogram according to octave scale. This option works only when have high SR recordings.
if (octaveScale)
{
// ASSUME fixed Occtave scale - USEFUL ONLY FOR JASCO 64000sr MARINE RECORDINGS
// If you wish to use other octave scale types then need to put in the config file and and set up recovery here.
dspOutput2.AmplitudeSpectrogram = OctaveFreqScale.AmplitudeSpectra(
dspOutput2.AmplitudeSpectrogram,
dspOutput2.WindowPower,
sampleRate,
epsilon,
freqScale);
dspOutput2.NyquistBin = dspOutput2.AmplitudeSpectrogram.GetLength(1) - 1; // ASSUMPTION!!! Nyquist is in top Octave bin - not necessarily true!!
// Calculate the octave scaled spectrogram.
dspOutput2.AmplitudeSpectrogram = OctaveFreqScale.ConvertLinearSpectrogramToOctaveFreqScale(dspOutput2.AmplitudeSpectrogram, freqScale);
dspOutput2.NyquistBin = dspOutput2.AmplitudeSpectrogram.GetLength(1) - 1; // ASSUME Nyquist is in top Octave bin.
}
}

Expand Down
Loading

0 comments on commit 5f26e3f

Please sign in to comment.