public void SpectralIndexOsc_Test() { { var sourceRecording = PathHelper.ResolveAsset("Recordings", "BAC2_20071008-085040.wav"); // 1. Create temp directory to store output if (!this.outputDirectory.Exists) { this.outputDirectory.Create(); } // 2. Get the spectral index var recordingSegment = new AudioRecording(sourceRecording.FullName); var frameLength = Oscillations2014.DefaultFrameLength; var sampleLength = Oscillations2014.DefaultSampleLength; var threshold = Oscillations2014.DefaultSensitivityThreshold; var spectralIndex = Oscillations2014.GetSpectralIndex_Osc(recordingSegment, frameLength, sampleLength, threshold); // 3. construct name of spectral index vector // SAVE THE OUTPUT if true // WARNING: this will overwrite fixtures var sourceName = Path.GetFileNameWithoutExtension(sourceRecording.Name); var stem = sourceName + ".SpectralIndex.OSC"; var expectedIndexPath = PathHelper.ResolveAsset("Oscillations2014", stem + ".EXPECTED.csv"); if (false) { // 4. Save spectral index vector to file //Csv.WriteToCsv(expectedIndexPath, spectralIndex); //Json.Serialise(expectedIndexPath, spectralIndex); Binary.Serialize(expectedIndexPath, spectralIndex); // 5. Get the vector as image and save as image file // no need to do tests on this image but it is useful to visualise output var expectedVectorImage = ImageTools.DrawVectorInColour(DataTools.reverseArray(spectralIndex), cellWidth: 10); var expectedImagePath = PathHelper.ResolveAsset("Oscillations2014", stem + ".png"); expectedVectorImage.Save(expectedImagePath.FullName, ImageFormat.Png); } // 6. Get the vector as image and save as image file // no need to do tests on this image but it is useful to compare with expected visual output var currentVectorImage = ImageTools.DrawVectorInColour(DataTools.reverseArray(spectralIndex), cellWidth: 10); var currentImagePath = Path.Combine(this.outputDirectory.FullName, stem + ".png"); currentVectorImage.Save(currentImagePath, ImageFormat.Png); // 7. Run test. Compare vectors // TODO this test fails when using CSV reader because the reader cuts out first element/line of the vector //var expectedVector = (double[])Csv.ReadFromCsv<double>(expectedIndexPath); var expectedVector = Binary.Deserialize <double[]>(expectedIndexPath); CollectionAssert.That.AreEqual(expectedVector, spectralIndex, 0.000001); } }
//////public static IndexCalculateResult Analysis( public static SpectralIndexValuesForContentDescription Analysis( AudioRecording recording, TimeSpan segmentOffsetTimeSpan, int sampleRateOfOriginalAudioFile, bool returnSonogramInfo = false) { // returnSonogramInfo = true; // if debugging double epsilon = recording.Epsilon; int sampleRate = recording.WavReader.SampleRate; //var segmentDuration = TimeSpan.FromSeconds(recording.WavReader.Time.TotalSeconds); var indexCalculationDuration = TimeSpan.FromSeconds(ContentSignatures.IndexCalculationDurationInSeconds); // Get FRAME parameters for the calculation of Acoustic Indices int frameSize = ContentSignatures.FrameSize; int frameStep = frameSize; // that is, windowOverlap = zero double frameStepDuration = frameStep / (double)sampleRate; // fraction of a second var frameStepTimeSpan = TimeSpan.FromTicks((long)(frameStepDuration * TimeSpan.TicksPerSecond)); // INITIALISE a RESULTS STRUCTURE TO return // initialize a result object in which to store SummaryIndexValues and SpectralIndexValues etc. var config = new IndexCalculateConfig(); // sets some default values int freqBinCount = frameSize / 2; var indexProperties = GetIndexProperties(); ////////var result = new IndexCalculateResult(freqBinCount, indexProperties, indexCalculationDuration, segmentOffsetTimeSpan, config); var spectralIndices = new SpectralIndexValuesForContentDescription(); ///////result.SummaryIndexValues = null; ///////SpectralIndexValues spectralIndices = result.SpectralIndexValues; // set up default spectrogram to return ///////result.Sg = returnSonogramInfo ? GetSonogram(recording, windowSize: 1024) : null; ///////result.Hits = null; ///////result.TrackScores = new List<Plot>(); // ################################## FINISHED SET-UP // ################################## NOW GET THE AMPLITUDE SPECTROGRAM // EXTRACT ENVELOPE and SPECTROGRAM FROM RECORDING SEGMENT // Note that the amplitude spectrogram has had the DC bin removed. i.e. has only 256 columns. var dspOutput1 = DSP_Frames.ExtractEnvelopeAndFfts(recording, frameSize, frameStep); var amplitudeSpectrogram = dspOutput1.AmplitudeSpectrogram; // (B) ################################## EXTRACT OSC SPECTRAL INDEX DIRECTLY FROM THE RECORDING ################################## // Get the oscillation spectral index OSC separately from signal because need a different frame size etc. var sampleLength = Oscillations2014.DefaultSampleLength; var frameLength = Oscillations2014.DefaultFrameLength; var sensitivity = Oscillations2014.DefaultSensitivityThreshold; var spectralIndexShort = Oscillations2014.GetSpectralIndex_Osc(recording, frameLength, sampleLength, sensitivity); // double length of the vector because want to work with 256 element vector for spectrogram purposes spectralIndices.OSC = DataTools.VectorDoubleLengthByAverageInterpolation(spectralIndexShort); // (C) ################################## EXTRACT SPECTRAL INDICES FROM THE AMPLITUDE SPECTROGRAM ################################## // IFF there has been UP-SAMPLING, calculate bin of the original audio nyquist. this will be less than SR/2. // original sample rate can be anything 11.0-44.1 kHz. int originalNyquist = sampleRateOfOriginalAudioFile / 2; // if up-sampling has been done if (dspOutput1.NyquistFreq > originalNyquist) { dspOutput1.NyquistFreq = originalNyquist; dspOutput1.NyquistBin = (int)Math.Floor(originalNyquist / dspOutput1.FreqBinWidth); // note that bin width does not change } // ii: CALCULATE THE ACOUSTIC COMPLEXITY INDEX spectralIndices.ACI = AcousticComplexityIndex.CalculateAci(amplitudeSpectrogram); // iii: CALCULATE the H(t) or Temporal ENTROPY Spectrum and then reverse the values i.e. calculate 1-Ht for energy concentration double[] temporalEntropySpectrum = AcousticEntropy.CalculateTemporalEntropySpectrum(amplitudeSpectrogram); for (int i = 0; i < temporalEntropySpectrum.Length; i++) { temporalEntropySpectrum[i] = 1 - temporalEntropySpectrum[i]; } spectralIndices.ENT = temporalEntropySpectrum; // (C) ################################## EXTRACT SPECTRAL INDICES FROM THE DECIBEL SPECTROGRAM ################################## // i: Convert amplitude spectrogram to decibels and calculate the dB background noise profile double[,] decibelSpectrogram = MFCCStuff.DecibelSpectra(dspOutput1.AmplitudeSpectrogram, dspOutput1.WindowPower, sampleRate, epsilon); double[] spectralDecibelBgn = NoiseProfile.CalculateBackgroundNoise(decibelSpectrogram); spectralIndices.BGN = spectralDecibelBgn; // ii: Calculate the noise reduced decibel spectrogram derived from segment recording. // REUSE the var decibelSpectrogram but this time using dspOutput1. decibelSpectrogram = MFCCStuff.DecibelSpectra(dspOutput1.AmplitudeSpectrogram, dspOutput1.WindowPower, sampleRate, epsilon); decibelSpectrogram = SNR.TruncateBgNoiseFromSpectrogram(decibelSpectrogram, spectralDecibelBgn); decibelSpectrogram = SNR.RemoveNeighbourhoodBackgroundNoise(decibelSpectrogram, nhThreshold: 2.0); // iii: CALCULATE noise reduced AVERAGE DECIBEL SPECTRUM spectralIndices.PMN = SpectrogramTools.CalculateAvgDecibelSpectrumFromDecibelSpectrogram(decibelSpectrogram); // ###################################################################################################################################################### // iv: CALCULATE SPECTRAL COVER. NOTE: at this point, decibelSpectrogram is noise reduced. All values >= 0.0 // FreqBinWidth can be accessed, if required, through dspOutput1.FreqBinWidth // dB THRESHOLD for calculating spectral coverage double dBThreshold = ActivityAndCover.DefaultActivityThresholdDb; // Calculate lower and upper boundary bin ids. // Boundary between low & mid frequency bands is to avoid low freq bins containing anthropogenic noise. These biased index values away from bio-phony. int midFreqBound = config.MidFreqBound; int lowFreqBound = config.LowFreqBound; int lowerBinBound = (int)Math.Ceiling(lowFreqBound / dspOutput1.FreqBinWidth); int middleBinBound = (int)Math.Ceiling(midFreqBound / dspOutput1.FreqBinWidth); var spActivity = ActivityAndCover.CalculateSpectralEvents(decibelSpectrogram, dBThreshold, frameStepTimeSpan, lowerBinBound, middleBinBound); //spectralIndices.CVR = spActivity.CoverSpectrum; spectralIndices.EVN = spActivity.EventSpectrum; ///////result.TrackScores = null; ///////return result; return(spectralIndices); } // end calculation of Six Spectral Indices
public static IndexCalculateResult Analysis( AudioRecording recording, TimeSpan subsegmentOffsetTimeSpan, Dictionary <string, IndexProperties> indexProperties, int sampleRateOfOriginalAudioFile, TimeSpan segmentStartOffset, IndexCalculateConfig config, bool returnSonogramInfo = false) { // returnSonogramInfo = true; // if debugging double epsilon = recording.Epsilon; int signalLength = recording.WavReader.GetChannel(0).Length; int sampleRate = recording.WavReader.SampleRate; var segmentDuration = TimeSpan.FromSeconds(recording.WavReader.Time.TotalSeconds); var indexCalculationDuration = config.IndexCalculationDurationTimeSpan; int nyquist = sampleRate / 2; // Get FRAME parameters for the calculation of Acoustic Indices //WARNING: DO NOT USE Frame Overlap when calculating acoustic indices. // It yields ACI, BGN, POW and EVN results that are significantly different from the default. // I have not had time to check if the difference is meaningful. Best to avoid. //int frameSize = (int?)config[AnalysisKeys.FrameLength] ?? IndexCalculateConfig.DefaultWindowSize; int frameSize = config.FrameLength; int frameStep = frameSize; // that is, windowOverlap = zero double frameStepDuration = frameStep / (double)sampleRate; // fraction of a second var frameStepTimeSpan = TimeSpan.FromTicks((long)(frameStepDuration * TimeSpan.TicksPerSecond)); int midFreqBound = config.MidFreqBound; int lowFreqBound = config.LowFreqBound; int freqBinCount = frameSize / 2; // double freqBinWidth = recording.Nyquist / (double)freqBinCount; // get duration in seconds and sample count and frame count double subsegmentDurationInSeconds = indexCalculationDuration.TotalSeconds; int subsegmentSampleCount = (int)(subsegmentDurationInSeconds * sampleRate); double subsegmentFrameCount = subsegmentSampleCount / (double)frameStep; subsegmentFrameCount = (int)Math.Ceiling(subsegmentFrameCount); // In order not to lose the last fractional frame, round up the frame number // and get the exact number of samples in the integer number of frames. // Do this because when IndexCalculationDuration = 100ms, the number of frames is only 8. subsegmentSampleCount = (int)(subsegmentFrameCount * frameStep); // get start and end samples of the subsegment and noise segment double localOffsetInSeconds = subsegmentOffsetTimeSpan.TotalSeconds - segmentStartOffset.TotalSeconds; int startSample = (int)(localOffsetInSeconds * sampleRate); int endSample = startSample + subsegmentSampleCount - 1; // Default behaviour: set SUBSEGMENT = total recording var subsegmentRecording = recording; // But if the indexCalculationDuration < segmentDuration if (indexCalculationDuration < segmentDuration) { // minimum samples needed to calculate acoustic indices. This value was chosen somewhat arbitrarily. // It allowes for case where IndexCalculationDuration = 100ms which is approx 8 frames int minimumViableSampleCount = frameSize * 8; int availableSignal = signalLength - startSample; // if (the required audio is beyond recording OR insufficient for analysis) then backtrack. if (availableSignal < minimumViableSampleCount) { // Back-track so we can fill a whole result. // This is a silent correction, equivalent to having a segment overlap for the last segment. var oldStart = startSample; startSample = signalLength - subsegmentSampleCount; endSample = signalLength; Logger.Trace(" Backtrack subsegment to fill missing data from imperfect audio cuts because not enough samples available. " + (oldStart - startSample) + " samples overlap."); } var subsamples = DataTools.Subarray(recording.WavReader.Samples, startSample, subsegmentSampleCount); var wr = new Acoustics.Tools.Wav.WavReader(subsamples, 1, 16, sampleRate); subsegmentRecording = new AudioRecording(wr); } // INITIALISE a RESULTS STRUCTURE TO return // initialize a result object in which to store SummaryIndexValues and SpectralIndexValues etc. var result = new IndexCalculateResult(freqBinCount, indexProperties, indexCalculationDuration, subsegmentOffsetTimeSpan, config); SummaryIndexValues summaryIndices = result.SummaryIndexValues; SpectralIndexValues spectralIndices = result.SpectralIndexValues; // set up default spectrogram to return result.Sg = returnSonogramInfo ? GetSonogram(recording, windowSize: 1024) : null; result.Hits = null; result.TrackScores = new List <Plot>(); // ################################## FINSIHED SET-UP // ################################## NOW GET THE AMPLITUDE SPECTORGRAMS // EXTRACT ENVELOPE and SPECTROGRAM FROM SUBSEGMENT var dspOutput1 = DSP_Frames.ExtractEnvelopeAndFfts(subsegmentRecording, frameSize, frameStep); // Select band according to min and max bandwidth int minBand = (int)(dspOutput1.AmplitudeSpectrogram.GetLength(1) * config.MinBandWidth); int maxBand = (int)(dspOutput1.AmplitudeSpectrogram.GetLength(1) * config.MaxBandWidth) - 1; dspOutput1.AmplitudeSpectrogram = MatrixTools.Submatrix( dspOutput1.AmplitudeSpectrogram, 0, minBand, dspOutput1.AmplitudeSpectrogram.GetLength(0) - 1, maxBand); // TODO: Michael to review whether bandwidth filter should be moved to DSP_Frames?? // Recalculate NyquistBin and FreqBinWidth, because they change with band selection //dspOutput1.NyquistBin = dspOutput1.AmplitudeSpectrogram.GetLength(1) - 1; //dspOutput1.FreqBinWidth = sampleRate / (double)dspOutput1.AmplitudeSpectrogram.GetLength(1) / 2; // Linear or Octave or Mel frequency scale? Set Linear as default. var freqScale = new FrequencyScale(nyquist: nyquist, frameSize: frameSize, hertzGridInterval: 1000); var freqScaleType = config.FrequencyScale; bool octaveScale = freqScaleType == FreqScaleType.Linear125Octaves7Tones28Nyquist32000; bool melScale = freqScaleType == FreqScaleType.Mel; if (octaveScale) { // only allow one octave scale at the moment - for Jasco marine recordings. // 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. freqScale = new FrequencyScale(FreqScaleType.Linear125Octaves7Tones28Nyquist32000); // 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!! } else if (melScale) { int minFreq = 0; int maxFreq = recording.Nyquist; dspOutput1.AmplitudeSpectrogram = MFCCStuff.MelFilterBank( dspOutput1.AmplitudeSpectrogram, config.MelScale, recording.Nyquist, minFreq, 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 // If the index calculation duration >= 30 seconds, then calculate BGN from the existing segment of recording. bool doSeparateBgnNoiseCalculation = indexCalculationDuration.TotalSeconds + (2 * config.BgNoiseBuffer.TotalSeconds) < segmentDuration.TotalSeconds / 2; var dspOutput2 = dspOutput1; if (doSeparateBgnNoiseCalculation) { // GET a longer SUBSEGMENT FOR NOISE calculation with 5 sec buffer on either side. // If the index calculation duration is shorter than 30 seconds, then need to calculate BGN noise from a longer length of recording // i.e. need to add noiseBuffer either side. Typical noiseBuffer value = 5 seconds int sampleBuffer = (int)(config.BgNoiseBuffer.TotalSeconds * sampleRate); var bgnRecording = AudioRecording.GetRecordingSubsegment(recording, startSample, endSample, sampleBuffer); // EXTRACT ENVELOPE and SPECTROGRAM FROM BACKGROUND NOISE SUBSEGMENT dspOutput2 = DSP_Frames.ExtractEnvelopeAndFfts(bgnRecording, frameSize, frameStep); // 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!! } } // ###################################### BEGIN CALCULATION OF INDICES ################################## // (A) ################################## EXTRACT SUMMARY INDICES FROM THE SIGNAL WAVEFORM ################################## // average absolute value over the minute recording - not useful // double[] avAbsolute = dspOutput1.Average; double[] signalEnvelope = dspOutput1.Envelope; double avgSignalEnvelope = signalEnvelope.Average(); // 10 times log of amplitude squared summaryIndices.AvgSignalAmplitude = 20 * Math.Log10(avgSignalEnvelope); // Deal with case where the signal waveform is continuous flat with values < 0.001. Has happened!! // Although signal appears zero, this condition is required. if (avgSignalEnvelope < 0.0001) { Logger.Debug("Segment skipped because avSignalEnvelope is < 0.001!"); summaryIndices.ZeroSignal = 1.0; return(result); } // i. Check for clipping and high amplitude rates per second summaryIndices.HighAmplitudeIndex = dspOutput1.HighAmplitudeCount / subsegmentDurationInSeconds; summaryIndices.ClippingIndex = dspOutput1.ClipCount / subsegmentDurationInSeconds; // ii. Calculate bg noise in dB // Convert signal envelope to dB and subtract background noise. Default noise SD to calculate threshold = ZERO double signalBgn = NoiseRemovalModal.CalculateBackgroundNoise(dspOutput2.Envelope); summaryIndices.BackgroundNoise = signalBgn; // iii: FRAME ENERGIES - convert signal to decibels and subtract background noise. double[] dBEnvelope = SNR.Signal2Decibels(dspOutput1.Envelope); double[] dBEnvelopeSansNoise = SNR.SubtractAndTruncate2Zero(dBEnvelope, signalBgn); // iv: ACTIVITY for NOISE REDUCED SIGNAL ENVELOPE // Calculate fraction of frames having acoustic activity var activity = ActivityAndCover.CalculateActivity(dBEnvelopeSansNoise, frameStepTimeSpan); summaryIndices.Activity = activity.FractionOfActiveFrames; // v. average number of events per second whose duration > one frame // average event duration in milliseconds - no longer calculated //summaryIndices.AvgEventDuration = activity.avEventDuration; summaryIndices.EventsPerSecond = activity.EventCount / subsegmentDurationInSeconds; // vi. Calculate SNR and active frames SNR summaryIndices.Snr = dBEnvelopeSansNoise.Max(); summaryIndices.AvgSnrOfActiveFrames = activity.ActiveAvDb; // vii. ENTROPY of ENERGY ENVELOPE -- 1-Ht because want measure of concentration of acoustic energy. double entropy = DataTools.EntropyNormalised(DataTools.SquareValues(signalEnvelope)); summaryIndices.TemporalEntropy = 1 - entropy; // Note that the spectrogram has had the DC bin removed. i.e. has only 256 columns. double[,] amplitudeSpectrogram = dspOutput1.AmplitudeSpectrogram; // get amplitude spectrogram. // CALCULATE various NDSI (Normalised difference soundscape Index) FROM THE AMPLITUDE SPECTROGRAM // These options proved to be highly correlated. Therefore only use tuple.Item 1 which derived from Power Spectral Density. var tuple3 = SpectrogramTools.CalculateAvgSpectrumAndVarianceSpectrumFromAmplitudeSpectrogram(amplitudeSpectrogram); summaryIndices.Ndsi = SpectrogramTools.CalculateNdsi(tuple3.Item1, sampleRate, 1000, 2000, 8000); // (B) ################################## EXTRACT OSC SPECTRAL INDEX DIRECTLY FROM THE RECORDING ################################## // Get the oscillation spectral index OSC separately from signal because need a different frame size etc. var sampleLength = Oscillations2014.DefaultSampleLength; var frameLength = Oscillations2014.DefaultFrameLength; var sensitivity = Oscillations2014.DefaultSensitivityThreshold; var spectralIndexShort = Oscillations2014.GetSpectralIndex_Osc(subsegmentRecording, frameLength, sampleLength, sensitivity); // double length of the vector because want to work with 256 element vector for LDFC purposes spectralIndices.OSC = DataTools.VectorDoubleLengthByAverageInterpolation(spectralIndexShort); // (C) ################################## EXTRACT SPECTRAL INDICES FROM THE AMPLITUDE SPECTROGRAM ################################## // i: CALCULATE SPECTRUM OF THE SUM OF FREQ BIN AMPLITUDES - used for later calculation of ACI spectralIndices.SUM = MatrixTools.SumColumns(amplitudeSpectrogram); // Calculate lower and upper boundary bin ids. // Boundary between low & mid frequency bands is to avoid low freq bins containing anthropogenic noise. These biased index values away from biophony. // Boundary of upper bird-band is to avoid high freq artefacts due to mp3. int lowerBinBound = (int)Math.Ceiling(lowFreqBound / dspOutput1.FreqBinWidth); int middleBinBound = (int)Math.Ceiling(midFreqBound / dspOutput1.FreqBinWidth); // calculate number of freq bins in the bird-band. int midBandBinCount = middleBinBound - lowerBinBound + 1; if (octaveScale) { // the above frequency bin bounds do not apply with octave scale. Need to recalculate them suitable for Octave scale recording. lowFreqBound = freqScale.LinearBound; lowerBinBound = freqScale.GetBinIdForHerzValue(lowFreqBound); midFreqBound = 8000; // This value appears suitable for Jasco Marine recordings. Not much happens above 8kHz. //middleBinBound = freqScale.GetBinIdForHerzValue(midFreqBound); middleBinBound = freqScale.GetBinIdInReducedSpectrogramForHerzValue(midFreqBound); midBandBinCount = middleBinBound - lowerBinBound + 1; } // IFF there has been UP-SAMPLING, calculate bin of the original audio nyquist. this will be less than SR/2. // original sample rate can be anything 11.0-44.1 kHz. int originalNyquist = sampleRateOfOriginalAudioFile / 2; // if upsampling has been done if (dspOutput1.NyquistFreq > originalNyquist) { dspOutput1.NyquistFreq = originalNyquist; dspOutput1.NyquistBin = (int)Math.Floor(originalNyquist / dspOutput1.FreqBinWidth); // note that binwidth does not change } // ii: CALCULATE THE ACOUSTIC COMPLEXITY INDEX spectralIndices.DIF = AcousticComplexityIndex.SumOfAmplitudeDifferences(amplitudeSpectrogram); double[] aciSpectrum = AcousticComplexityIndex.CalculateAci(amplitudeSpectrogram); spectralIndices.ACI = aciSpectrum; // remove low freq band of ACI spectrum and store average ACI value double[] reducedAciSpectrum = DataTools.Subarray(aciSpectrum, lowerBinBound, midBandBinCount); summaryIndices.AcousticComplexity = reducedAciSpectrum.Average(); // iii: CALCULATE the H(t) or Temporal ENTROPY Spectrum and then reverse the values i.e. calculate 1-Ht for energy concentration double[] temporalEntropySpectrum = AcousticEntropy.CalculateTemporalEntropySpectrum(amplitudeSpectrogram); for (int i = 0; i < temporalEntropySpectrum.Length; i++) { temporalEntropySpectrum[i] = 1 - temporalEntropySpectrum[i]; } spectralIndices.ENT = temporalEntropySpectrum; // iv: remove background noise from the amplitude spectrogram // First calculate the noise profile from the amplitude sepctrogram double[] spectralAmplitudeBgn = NoiseProfile.CalculateBackgroundNoise(dspOutput2.AmplitudeSpectrogram); amplitudeSpectrogram = SNR.TruncateBgNoiseFromSpectrogram(amplitudeSpectrogram, spectralAmplitudeBgn); // AMPLITUDE THRESHOLD for smoothing background, nhThreshold, assumes background noise ranges around -40dB. // This value corresponds to approximately 6dB above backgorund. amplitudeSpectrogram = SNR.RemoveNeighbourhoodBackgroundNoise(amplitudeSpectrogram, nhThreshold: 0.015); ////ImageTools.DrawMatrix(spectrogramData, @"C:\SensorNetworks\WavFiles\Crows\image.png", false); ////DataTools.writeBarGraph(modalValues); result.AmplitudeSpectrogram = amplitudeSpectrogram; // v: ENTROPY OF AVERAGE SPECTRUM & VARIANCE SPECTRUM - at this point the spectrogram is a noise reduced amplitude spectrogram var tuple = AcousticEntropy.CalculateSpectralEntropies(amplitudeSpectrogram, lowerBinBound, midBandBinCount); // ENTROPY of spectral averages - Reverse the values i.e. calculate 1-Hs and 1-Hv, and 1-Hcov for energy concentration summaryIndices.EntropyOfAverageSpectrum = 1 - tuple.Item1; // ENTROPY of spectrum of Variance values summaryIndices.EntropyOfVarianceSpectrum = 1 - tuple.Item2; // ENTROPY of spectrum of Coefficient of Variation values summaryIndices.EntropyOfCoVSpectrum = 1 - tuple.Item3; // vi: ENTROPY OF DISTRIBUTION of maximum SPECTRAL PEAKS. // First extract High band SPECTROGRAM which is now noise reduced double entropyOfPeaksSpectrum = AcousticEntropy.CalculateEntropyOfSpectralPeaks(amplitudeSpectrogram, lowerBinBound, middleBinBound); summaryIndices.EntropyOfPeaksSpectrum = 1 - entropyOfPeaksSpectrum; // ###################################################################################################################################################### // (C) ################################## EXTRACT SPECTRAL INDICES FROM THE DECIBEL SPECTROGRAM ################################## // i: Convert amplitude spectrogram to deciBels and calculate the dB background noise profile double[,] deciBelSpectrogram = MFCCStuff.DecibelSpectra(dspOutput2.AmplitudeSpectrogram, dspOutput2.WindowPower, sampleRate, epsilon); double[] spectralDecibelBgn = NoiseProfile.CalculateBackgroundNoise(deciBelSpectrogram); spectralIndices.BGN = spectralDecibelBgn; // ii: Calculate the noise reduced decibel spectrogram derived from segment recording. // REUSE the var decibelSpectrogram but this time using dspOutput1. deciBelSpectrogram = MFCCStuff.DecibelSpectra(dspOutput1.AmplitudeSpectrogram, dspOutput1.WindowPower, sampleRate, epsilon); deciBelSpectrogram = SNR.TruncateBgNoiseFromSpectrogram(deciBelSpectrogram, spectralDecibelBgn); deciBelSpectrogram = SNR.RemoveNeighbourhoodBackgroundNoise(deciBelSpectrogram, nhThreshold: 2.0); // iii: CALCULATE noise reduced AVERAGE DECIBEL SPECTRUM spectralIndices.PMN = SpectrogramTools.CalculateAvgDecibelSpectrumFromDecibelSpectrogram(deciBelSpectrogram); // iv: CALCULATE SPECTRAL COVER. // NOTE: at this point, decibelSpectrogram is noise reduced. All values >= 0.0 // FreqBinWidth can be accessed, if required, through dspOutput1.FreqBinWidth double dBThreshold = ActivityAndCover.DefaultActivityThresholdDb; // dB THRESHOLD for calculating spectral coverage var spActivity = ActivityAndCover.CalculateSpectralEvents(deciBelSpectrogram, dBThreshold, frameStepTimeSpan, lowerBinBound, middleBinBound); spectralIndices.CVR = spActivity.CoverSpectrum; spectralIndices.EVN = spActivity.EventSpectrum; summaryIndices.HighFreqCover = spActivity.HighFreqBandCover; summaryIndices.MidFreqCover = spActivity.MidFreqBandCover; summaryIndices.LowFreqCover = spActivity.LowFreqBandCover; // ###################################################################################################################################################### // v: CALCULATE SPECTRAL PEAK TRACKS and RIDGE indices. // NOTE: at this point, the var decibelSpectrogram is noise reduced. i.e. all its values >= 0.0 // Detecting ridges or spectral peak tracks requires using a 5x5 mask which has edge effects. // This becomes significant if we have a short indexCalculationDuration. // Consequently if the indexCalculationDuration < 10 seconds then we revert back to the recording and cut out a recording segment that includes // a buffer for edge effects. In most cases however, we can just use the decibel spectrogram already calculated and ignore the edge effects. double peakThreshold = 6.0; //dB SpectralPeakTracks sptInfo; if (indexCalculationDuration.TotalSeconds < 10.0) { // calculate a new decibel spectrogram sptInfo = SpectralPeakTracks.CalculateSpectralPeakTracks(recording, startSample, endSample, frameSize, octaveScale, peakThreshold); } else { // use existing decibel spectrogram sptInfo = new SpectralPeakTracks(deciBelSpectrogram, peakThreshold); } spectralIndices.SPT = sptInfo.SptSpectrum; spectralIndices.RHZ = sptInfo.RhzSpectrum; spectralIndices.RVT = sptInfo.RvtSpectrum; spectralIndices.RPS = sptInfo.RpsSpectrum; spectralIndices.RNG = sptInfo.RngSpectrum; summaryIndices.SptDensity = sptInfo.TrackDensity; // these are two other indices that I tried but they do not seem to add anything of interest. //summaryIndices.AvgSptDuration = sptInfo.AvTrackDuration; //summaryIndices.SptPerSecond = sptInfo.TotalTrackCount / subsegmentSecondsDuration; // ###################################################################################################################################################### // vi: CLUSTERING - FIRST DETERMINE IF IT IS WORTH DOING // return if (activeFrameCount too small || eventCount == 0 || short index calc duration) because no point doing clustering if (activity.ActiveFrameCount <= 2 || Math.Abs(activity.EventCount) < 0.01 || indexCalculationDuration.TotalSeconds < 15) { // IN ADDITION return if indexCalculationDuration < 15 seconds because no point doing clustering on short time segment // NOTE: Activity was calculated with 3dB threshold AFTER backgroundnoise removal. //summaryIndices.AvgClusterDuration = TimeSpan.Zero; summaryIndices.ClusterCount = 0; summaryIndices.ThreeGramCount = 0; return(result); } // YES WE WILL DO CLUSTERING! to determine cluster count (spectral diversity) and spectral persistence. // Only use midband decibel SPECTRUM. In June 2016, the mid-band (i.e. the bird-band) was set to lowerBound=1000Hz, upperBound=8000hz. // Actually do clustering of binary spectra. Must first threshold double binaryThreshold = SpectralClustering.DefaultBinaryThresholdInDecibels; var midBandSpectrogram = MatrixTools.Submatrix(deciBelSpectrogram, 0, lowerBinBound, deciBelSpectrogram.GetLength(0) - 1, middleBinBound); var clusterInfo = SpectralClustering.ClusterTheSpectra(midBandSpectrogram, lowerBinBound, middleBinBound, binaryThreshold); // Store two summary index values from cluster info summaryIndices.ClusterCount = clusterInfo.ClusterCount; summaryIndices.ThreeGramCount = clusterInfo.TriGramUniqueCount; // As of May 2017, no longer store clustering results superimposed on spectrogram. // If you want to see this, then call the TEST methods in class SpectralClustering.cs. // ####################################################################################################################################################### // vii: set up other info to return var freqPeaks = SpectralPeakTracks.ConvertSpectralPeaksToNormalisedArray(deciBelSpectrogram); var scores = new List <Plot> { new Plot("Decibels", DataTools.normalise(dBEnvelopeSansNoise), ActivityAndCover.DefaultActivityThresholdDb), new Plot("Active Frames", DataTools.Bool2Binary(activity.ActiveFrames), 0.0), new Plot("Max Frequency", freqPeaks, 0.0), // relative location of freq maxima in spectra }; result.Hits = sptInfo.Peaks; result.TrackScores = scores; return(result); } // end Calculation of Summary and Spectral Indices