public InverseSTFT(IAudioWriterStream stream, int windowSize, int hopSize, int fftSize, WindowType windowType, float windowNormalizationFactor)
: base(stream, windowSize, hopSize)
{
if (fftSize < windowSize)
{
throw new ArgumentOutOfRangeException("fftSize must be >= windowSize");
}
frameBuffer = new float[fftSize];
fft = new PFFFT.PFFFT(fftSize, PFFFT.Transform.Real);
synthesisWindow = WindowUtil.GetFunction(windowType, windowSize, windowNormalizationFactor);
}
/// <summary>
/// Initializes a new STFT for the specified stream with the specified window, hop, and FFT size.
/// An FFT size larger than the window size can be used to increase frequency resolution. Window will be right-padded with zeros for FFT.
/// </summary>
/// <param name="stream">the stream to read the audio data to process from</param>
/// <param name="windowSize">the window size in the dimension of samples</param>
/// <param name="hopSize">the hop size in the dimension of samples</param>
/// <param name="fftSize">the FFT size, must be >= windowSize</param>
/// <param name="windowType">the type of the window function to apply</param>
/// <param name="outputFormat">format of the output data, e.g. raw FFT complex numbers or dB spectrum</param>
public STFT(IAudioStream stream, int windowSize, int hopSize, int fftSize, WindowType windowType, OutputFormat outputFormat)
: base(stream, windowSize, hopSize, windowType)
{
if (fftSize < windowSize)
{
throw new ArgumentOutOfRangeException("fftSize must be >= windowSize");
}
frameBuffer = new float[fftSize];
fftBuffer = new float[fftSize];
Array.Clear(frameBuffer, 0, frameBuffer.Length); // init with zeros (assure zero padding)
fft = new PFFFT.PFFFT(fftSize, PFFFT.Transform.Real);
this.outputFormat = outputFormat;
}
public static double FrequencyDistribution(byte[] x, byte[] y)
{
if (x.Length != y.Length)
{
throw new ArgumentException("interval lengths do not match");
}
int samples = x.Length / 4;
float[] xF = new float[samples];
float[] yF = new float[samples];
Buffer.BlockCopy(x, 0, xF, 0, x.Length);
Buffer.BlockCopy(y, 0, yF, 0, y.Length);
int fftSize = 2048; // max FFT size
// if there are less samples to analyze, find a fitting FFT size
while (fftSize > samples)
{
fftSize /= 2;
}
int hopSize = fftSize / 8;
if (fftSize < 64)
{
throw new Exception("FFT might be too small to get a meaningful result");
}
double[] xSum = new double[fftSize / 2];
double[] ySum = new double[fftSize / 2];
float[] buffer = new float[fftSize];
if (windowFunction == null || windowFunction.Size != fftSize)
{
windowFunction = WindowUtil.GetFunction(WindowType.Hann, fftSize);
}
var fft = new PFFFT.PFFFT(fftSize, PFFFT.Transform.Real);
int blocks = (samples - fftSize) / hopSize;
for (int block = 0; block < blocks; block++)
{
Array.Copy(xF, block * hopSize, buffer, 0, buffer.Length);
FrequencyDistributionBlockProcess(buffer, xSum, fft);
Array.Copy(yF, block * hopSize, buffer, 0, buffer.Length);
FrequencyDistributionBlockProcess(buffer, ySum, fft);
}
fft.Dispose();
// remove DC offset
xSum[0] = 0;
ySum[0] = 0;
double xScalarSum = xSum.Sum();
double yScalarSum = ySum.Sum();
double result = 0;
for (int i = 0; i < fftSize / 2; i++)
{
result += Math.Abs(xSum[i] / xScalarSum - ySum[i] / yScalarSum);
}
return(1 - result);
}
private static void FrequencyDistributionBlockProcess(float[] buffer, double[] target, PFFFT.PFFFT fft)
{
windowFunction.Apply(buffer);
fft.Forward(buffer);
for (int i = 0; i < buffer.Length; i += 2)
{
buffer[i / 2] = FFTUtil.CalculateMagnitude(buffer[i], buffer[i + 1]) / buffer.Length * 2;
}
for (int i = 0; i < target.Length; i++)
{
target[i] += buffer[i];
}
}
private void Generate(FFTLibrary fftLib)
{
// FFT resources:
// http://www.mathworks.de/support/tech-notes/1700/1702.html
// http://stackoverflow.com/questions/6627288/audio-spectrum-analysis-using-fft-algorithm-in-java
// http://stackoverflow.com/questions/1270018/explain-the-fft-to-me
// http://stackoverflow.com/questions/6666807/how-to-scale-fft-output-of-wave-file
/*
* FFT max value test:
* 16Hz, 1024 samplerate, 512 samples, rectangle window
* -> input min/max: -1/1
* -> FFT max: 256
* -> FFT result max: ~1
* -> FFT dB result max: ~1
*
* source: "Windowing Functions Improve FFT Results, Part I"
* => kann zum Normalisieren für Fensterfunktionen verwendet werden
*/
float frequency = float.Parse(frequencyTextBox.Text);
int ws = (int)windowSize.Value;
float frequencyFactor = ws / frequency;
int sampleRate = int.Parse(sampleRateTextBox.Text);
SineGeneratorStream sine = new SineGeneratorStream(sampleRate, frequency, new TimeSpan(0, 0, 1));
float[] input = new float[ws];
sine.Read(input, 0, ws);
//// add second frequency
//SineGeneratorStream sine2 = new SineGeneratorStream(44100, 700, new TimeSpan(0, 0, 1));
//float[] input2 = new float[ws];
//sine2.Read(input2, 0, ws);
//for (int x = 0; x < input.Length; x++) {
// input[x] += input2[x];
// input[x] /= 2;
//}
//// add dc offset
//for (int x = 0; x < input.Length; x++) {
// input[x] += 0.5f;
//}
inputGraph.Values = input;
WindowFunction wf = WindowUtil.GetFunction((WindowType)windowTypes.SelectedValue, ws);
float[] window = (float[])input.Clone();
wf.Apply(window);
input2Graph.Values = window;
float[] fftIO = (float[])window.Clone();
if (fftLib == FFTLibrary.ExocortexDSP)
{
// This function call indirection is needed to allow this method execute even when the
// Exocortex FFT assembly is missing.
ApplyExocortexDSP(fftIO);
}
else if (fftLib == FFTLibrary.FFTW)
{
FFTW.FFTW fftw = new FFTW.FFTW(fftIO.Length);
fftw.Execute(fftIO);
}
else if (fftLib == FFTLibrary.PFFFT)
{
PFFFT.PFFFT pffft = new PFFFT.PFFFT(fftIO.Length, PFFFT.Transform.Real);
pffft.Forward(fftIO, fftIO);
}
//// convert real input to complex input with im part set to zero
//float[] fftIO = new float[ws * 2];
//int i = 0;
//for (int j = 0; j < window.Length; j++) {
// fftIO[i] = window[j];
// i += 2;
//}
//Fourier.FFT(fftIO, fftIO.Length / 2, FourierDirection.Forward);
fftOutputGraph.Values = fftIO;
float[] fftResult = new float[ws / 2];
//FFTUtil.Results(fftIO, fftResult);
// transform complex output to magnitudes
float sum = 0;
int y = 0;
for (int x = 0; x < fftIO.Length; x += 2) // / 2
{
fftResult[y] = FFTUtil.CalculateMagnitude(fftIO[x], fftIO[x + 1]) / ws * 2;
sum += fftResult[y++];
}
//FFTUtil.Results(fftIO, fftResult);
//// adjust values for the sum to become 1
//float sum2 = 0;
//for (int x = 0; x < fftResult.Length; x++) {
// fftResult[x] /= sum;
// sum2 += fftResult[x];
//}
//Debug.WriteLine("sum / sum2: {0} / {1}", sum, sum2);
fftNormalizedOutputGraph.Values = fftResult;
// convert magnitudes to decibel
float[] fftResultdB = new float[fftResult.Length];
//for (int x = 0; x < fftResult.Length; x++) {
// fftResultdB[x] = (float)VolumeUtil.LinearToDecibel(fftResult[x]);
//}
FFTUtil.Results(fftIO, fftResultdB);
fftdBOutputGraph.Values = fftResultdB;
summary.Text = String.Format(
"input min/max: {0}/{1} -> window min/max: {2}/{3} -> fft min/max: {4}/{5} -> result min/max/bins/sum: {6}/{7}/{8}/{9} -> dB min/max: {10:0.000000}/{11:0.000000}",
input.Min(), input.Max(),
window.Min(), window.Max(),
fftIO.Min(), fftIO.Max(),
fftResult.Min(), fftResult.Max(), fftResult.Length, sum,
fftResultdB.Min(), fftResultdB.Max());
}