/// <summary>
/// Transforms one window of MFCCs. The following steps are
/// performed: <br>
/// <br>
/// (1) normalized power fft with hanning window function<br>
/// (2) convert to Mel scale by applying a mel filter bank<br>
/// (3) convertion to db<br>
/// (4) finally a DCT is performed to get the mfcc<br>
///<br>
/// This process is mathematical identical with the process described in [1].
/// </summary>
/// <param name="window">double[] data to be converted, must contain enough data for
/// one window</param>
/// <param name="start">int start index of the window data</param>
/// <returns>double[] the window representation in Sone</returns>
public double[] ProcessWindow(double[] window, int start)
{
//number of unique coefficients, and the rest are symmetrically redundant
int fftSize = (windowSize / 2) + 1;
//check start
if(start < 0)
throw new Exception("start must be a positive value");
//check window size
if(window == null || window.Length - start < windowSize)
throw new Exception("the given data array must not be a null value and must contain data for one window");
//just copy to buffer
for (int j = 0; j < windowSize; j++)
buffer[j] = window[j + start];
//perform power fft
normalizedPowerFFT.Transform(buffer, null);
//use all coefficient up to the nequist frequency (ceil((fftSize+1)/2))
Matrix x = new Matrix(buffer, windowSize);
x = x.GetMatrix(0, fftSize-1, 0, 0); //fftSize-1 is the index of the nyquist frequency
//apply mel filter banks
x = melFilterBanks.Times(x);
//to db
double log10 = 10 * (1 / Math.Log(10)); // log for base 10 and scale by factor 10
x.ThrunkAtLowerBoundary(1);
x.LogEquals();
x.TimesEquals(log10);
//compute DCT
x = dctMatrix.Times(x);
return x.GetColumnPackedCopy();
}
/// <summary>
/// Apply internal DCT and Mel Filterbands utilising the Comirva Matrix methods
/// </summary>
/// <param name="m">matrix (stftdata)</param>
/// <returns>matrix mel scaled and dct'ed</returns>
public Matrix ApplyMelScaleDCTComirva(ref Matrix m)
{
Mirage.DbgTimer t = new Mirage.DbgTimer();
t.Start();
// 4. Mel Scale Filterbank
// Mel-frequency is proportional to the logarithm of the linear frequency,
// reflecting similar effects in the human's subjective aural perception)
m = filterWeights * m;
// 5. Take Logarithm
// to db
double log10 = 20 * (1 / Math.Log(10)); // log for base 10 and scale by factor 10
m.ThrunkAtLowerBoundary(1);
m.LogEquals();
m *= log10;
// 6. DCT (Discrete Cosine Transform)
m = dct * m;
Mirage.Dbg.WriteLine("mfcc (MfccMirage-ComirvaWay) Execution Time: " + t.Stop().TotalMilliseconds + " ms");
return m;
}