Known bugs and changelog

Version 0.9.5

• Add methods for fractional delay FIR filter design
• Add code for designing even-sized FIR filters
• [planned] Add 3D audio functionality (stereo pan, HRTF)
• [planned] Add FadeIn/FadeOut functionality in signal builders
• [planned] Use Span in certain parts of the code (should test it carefully)
• [planned] Multi-target package to .NET Standard 2.0 and 2.1 (mainly, because of FastSpan) / .NET 5
• [planned] NWaves.Viz (separate package for cross-platform visualization / exporting to bmp files)
• Various minor enhancements / bugfixes:
  • implement 'perfect reconstruction' mode in inverse STFT
  • add normalization flag in Stft.Spectrogram() method
  • add DiscreteSignal.Unit() and DiscreteSignal.Reverse() methods
  • add EstimateGain() method and methods for filter processing with user-specified gain
  • WaveFile constructor doesn't close underlying stream
  • change type of TF zeros and poles to Complex[]. Methods accepting ComplexDiscreteSignal objects are not removed.
  • add overloaded constructor: var zsignal = new ComplexDiscreteSignal(1, tf.Zeros);

Version 0.9.4

• Add ZiFilter class / TransferFunction.Zi property
• Add methods for TransferFunction <-> StateSpace conversion
• Modify DesignFilter.IirBsTf() function to produce MATLAB/sciPy-like results
• Add FastDct family (DCT implementations via FFT for large sizes)
• Add Mdct and FastMdct classes (Modified DCT)
• Add KarplusStrongBuilder / KarplusStrongDrumBuilder classes
• Add PadSynthBuilder class
• Add StereoFilter class for convenient filtering of data in interleaved buffers
• Add Scale.NoteToFreq() / Scale.FreqToNote() methods
• Add Fft.InverseNorm() and RealFft.InverseNorm() methods
• Bugfix: correct gain in TfToSos() method
• Bugfix: correct results of RealFft.Inverse() (results of previous implementation should be multiplied by 2)

Version 0.9.3

• Accelerate FIR/IIR online filtering (up to ~2x in .NET Framework, up to ~1.3x in .NET Core)
• Add 64-bit FIR/IIR filters and block convolvers (OLA/OLS)
• Fix incorrect implementation of adaptive filtering
• Add OverlapAddFilter base class for PV-based filters/effects
• Add ByteConverter class
• Bugfix: allow STFT have window size different from power-of-two
• Introduce LogEnergyFloor parameter in MFCC/PNCC/PLP extractors
• Add chorus effect; enhance flanger, vibrato effects
• Nicer design of feature extractors:

var mfccOptions = new MfccOptions
{
    SamplingRate = 22050,
    FeatureCount = 13,
    FrameDuration = 0.032/*sec*/,
    HopDuration = 0.015/*sec*/,
    FilterBankSize = 26,
    HighFrequency = 6000,
    PreEmphasis = 0.97,
    //...unspecified parameters will have default values 
};

var mfccExtractor = new MfccExtractor(mfccOptions);
List<float[]> mfccVectors = mfccExtractor.ComputeFrom(signal);

// { 0, 0.015, 0.030, 0.045, 0.060, 0.075, ... }
List<double> timeMarkers = mfccExtractor.TimeMarkers(mfccVectors.Count);

// serialize current config to JSON file:
using (var config = new FileStream("file.json", FileMode.Create))
{
    config.SaveOptions(mfccOptions);
}

// open config from JSON file:
PlpOptions options;
using (var config = new FileStream("file.json", FileMode.Open))
{
    // identical fields will be copied, other fields will be ignored
    options = config.LoadOptions<PlpOptions>();
}

// cast (identical fields will be copied, other fields will be ignored)
options = mfccOptions.Cast<MfccOptions, PlpOptions>();

// get all errors after validation:
List<string> errors = options.Errors;

Version 0.9.2

• Add more IIR filters: Bessel, Chebyshev-I & II, Elliptic, Thiran
• Add TfToSos / SosToTf methods to DesignFilter class for working with second-order sections
• Add SavitzkyGolayFilter class (window size up to 31; derivatives 0, 1, 2)
• Add Fwt class for Fast Wavelet Transform and wavelets ("haar", "db1-20", "sym2-20", "coif1-5")
• Add PolyphaseSystem class
• Add PlpExtractor class
• Add Lpc class with new methods ToLsf / FromLsf, move LPC-related functions to Lpc
• Add more customizations in MfccExtractor class
• Add more customizations in FilterBanks class
• Add VtlnWarper class
• Add Stft methods: MagnitudePhaseSpectrogram()/ReconstructMagnitudePhase()
• Add HarmonicPercussiveSeparator class
• Add GriffinLimReconstructor class
• Add Remez class for equiripple FIR filter design
• Add methods FirWin(Lp|Hp|Bp|Bs) and FirEquiripple(Lp|Hp|Bp|Bs) to DesignFilter class
• Fix bug with pre-emphasis in PnccExtractor that led to less accurate results
• Add lpcOrder parameter to LpccExtractor
• Add RealFft class and use it instead of Fft where possible (25-35% performance gain)
• Important change in interface and implementation of LTI filter classes:

// code in versions before 0.9.2
var filter = DesignFilter.FirLp(145, 0.15f);
var impulseResponse = filter.ImpulseResponse();
var magnitudeResponse = filter.FrequencyResponse().Magnitude;
var tf = filter.Tf;

// became:
var filter = new FirFilter(DesignFilter.FirWinLp(145, 0.15));
var impulseResponse = filter.Tf.ImpulseResponse();
var magnitudeResponse = filter.Tf.FrequencyResponse().Magnitude;
var tf = filter.Tf;  // looks the same, but now it's more optimized

// in short:
// FirFilter / IirFilter classes are now used only for filtering.
// Everything related to filter design & analysis as of ver.0.9.2
// is concentrated in TransferFunction class. It was here before,
// but FDA responsibilities were somewhat mixed between filters and Tf.

// Filter objects contained arrays of both 32-bit and 64-bit coefficients
// although filtering was (and still is) done with 32-bit floats
// and FDA was (and still is) done with 64-bit doubles.

// As of ver.0.9.2 filters contain only 32-bit coefficients for filtering
// and the reference to TransferFunction object (not null - only if needed).
// TF objects contain 64-bit numerator / denominator coefficients.

var filter = new IirFilter(new[] { 1, 0.2, -0.3 }, new[] { 1, 0.9 });
var tf = filter.Tf;
// in this case TransferFunction will be generated on the fly
// from 32-bit floats (so we'll lose precision).
// Internal Tf object of the filter is still null.
// Memory efficient, but this filter maybe lacks precision for filter analysis.
// If we need just to filter data, this is the best solution.

var filter = new IirFilter(new TransferFunction(new[]{ 1, 0.2, -0.3 }, new[]{ 1, 0.9 }));
var tf = filter.Tf;
// here the filter stores the reference to TF with 64-bit precision
// and Tf property returns this reference.
// If we use this filter only for filtering, then this constructor is redundant
// and memory inefficient.


// Surely, we can work just with TF object, without any filter objects:

var tf = new TransferFunction(new[] { 1, 0.2, -0.3 }, new[] { 1, 0.9 });
var zeros = tf.Zeros;
var freqResponse = tf.FrequencyResponse(1024);
// etc.

// 64-bit precision:
var tf1 = new TransferFunction(b, a);
var tf2 = new TransferFunction(d, c);
var tf = tf1 * tf2;

// 32-bit precision:
var filter1 = new IirFilter(b, a);
var filter2 = new IirFilter(d, c);
var filter = filter1 * filter2;


// also:
var hpFilter = DesignFilter.LpToHp(filter);
var lpFilter = DesignFilter.HpToLp(hpFilter);
var bpFilter = DesignFilter.FirBp(123, 0.05f, 0.15f);
var brFilter = DesignFilter.FirBr(201, 0.08f, 0.23f, WindowTypes.Kaiser);

// became:
var hpFilter = DesignFilter.FirLpToHp(filter);
var lpFilter = DesignFilter.FirHpToLp(hpFilter);
var bpFilter = DesignFilter.FirWinBp(123, 0.05, 0.15);
var brFilter = DesignFilter.FirWinBr(201, 0.08, 0.23, WindowTypes.Kaiser);

Version 0.9.1

• Add adaptive filters (LMS variations, LMF, RLS)
• Add Hartley transform
• Add Mellin transform
• Add wavetable and ADSR signal builders
• Add PitchShiftVocoderEffect class
• Add parameter int parallelThreads in ParallelComputeFrom method of feature extractors
• Add overloaded versions of ParallelComputeFrom methods with startSample and endSample parameters
• Implement online processing methods in SpectralSubtractor, WhisperEffect, RobotEffect, MorphEffect
• Phase Vocoder with Identity Phase Locking is used by default for TSM (instead of WSOLA)
• Trim feature names in FeatureDescriptions list of multi-feature extractors
• Move PitchExtractor to FeatureExtractors namespace
• Fix bug with incorrectly working pre-emphasis filter in Lpc/LpccExtractor
• Size of FFT in Pncc/SpnccExtractor is 0 by default (it'll be auto-derived)