cannam@26: /* -*- c-basic-offset: 4 indent-tabs-mode: nil -*-  vi:set ts=8 sts=4 sw=4: */
cannam@25: 
cannam@25: /*
cannam@26:     QM DSP Library
cannam@25: 
cannam@26:     Centre for Digital Music, Queen Mary, University of London.
cannam@26:     This file copyright 2005 Nicolas Chetry, copyright 2008 QMUL.
cannam@26:     All rights reserved.
cannam@26: */
cannam@25: 
cannam@26: #include <cmath>
cannam@26: #include <cstdlib>
cannam@26: 
cannam@26: #include "MFCC.h"
cannam@26: #include "dsp/transforms/FFT.h"
cannam@26: #include "base/Window.h"
cannam@26: 
cannam@26: MFCC::MFCC(MFCCConfig config)
cannam@26: {
cannam@26:     int i,j;
cannam@26: 
cannam@26:     /* Calculate at startup */
cannam@26:     double *freqs, *lower, *center, *upper, *triangleHeight, *fftFreqs;
cannam@25:   
cannam@26:     lowestFrequency   = 66.6666666;
cannam@26:     linearFilters     = 13;
cannam@26:     linearSpacing     = 66.66666666;
cannam@26:     logFilters        = 27;
cannam@26:     logSpacing        = 1.0711703;
cannam@25:   
cannam@26:     /* FFT and analysis window sizes */
cannam@26:     fftSize           = config.fftsize;
cannam@26: 
cannam@26:     totalFilters      = linearFilters + logFilters;
cannam@25:   
cannam@26:     samplingRate      = config.FS;
cannam@26:   
cannam@26:     /* The number of cepstral componenents */
cannam@26:     nceps             = config.nceps;
cannam@25: 
cannam@26:     /* Set if user want C0 */
cannam@26:     WANT_C0           = (config.want_c0 ? 1 : 0);
cannam@25:   
cannam@26:     /* Allocate space for feature vector */
cannam@26:     if (WANT_C0 == 1) {
cannam@26:         ceps              = (double*)calloc(nceps+1, sizeof(double));
cannam@26:     } else {
cannam@26:         ceps              = (double*)calloc(nceps, sizeof(double));
cannam@26:     }
cannam@26:  
cannam@26:     /* Allocate space for local vectors */
cannam@26:     mfccDCTMatrix     = (double**)calloc(nceps+1, sizeof(double*));
cannam@26:     for (i=0;i<nceps+1; i++) {
cannam@26:         mfccDCTMatrix[i]= (double*)calloc(totalFilters, sizeof(double)); 
cannam@26:     }
cannam@26: 
cannam@26:     mfccFilterWeights = (double**)calloc(totalFilters, sizeof(double*));
cannam@26:     for (i=0;i<totalFilters; i++) {
cannam@26:         mfccFilterWeights[i] = (double*)calloc(fftSize, sizeof(double)); 
cannam@26:     }
cannam@26:     
cannam@26:     freqs  = (double*)calloc(totalFilters+2,sizeof(double));
cannam@26:     
cannam@26:     lower  = (double*)calloc(totalFilters,sizeof(double));
cannam@26:     center = (double*)calloc(totalFilters,sizeof(double));
cannam@26:     upper  = (double*)calloc(totalFilters,sizeof(double));
cannam@26:     
cannam@26:     triangleHeight = (double*)calloc(totalFilters,sizeof(double));
cannam@26:     fftFreqs       = (double*)calloc(fftSize,sizeof(double));
cannam@25:   
cannam@26:     for (i=0;i<linearFilters;i++) {
cannam@26:         freqs[i] = lowestFrequency + ((double)i) * linearSpacing;
cannam@26:     }
cannam@26:   
cannam@26:     for (i=linearFilters; i<totalFilters+2; i++) {
cannam@26:         freqs[i] = freqs[linearFilters-1] * 
cannam@26:             pow(logSpacing, (double)(i-linearFilters+1));
cannam@26:     }
cannam@26:   
cannam@26:     /* Define lower, center and upper */
cannam@26:     memcpy(lower,  freqs,totalFilters*sizeof(double));
cannam@26:     memcpy(center, &freqs[1],totalFilters*sizeof(double));
cannam@26:     memcpy(upper,  &freqs[2],totalFilters*sizeof(double));
cannam@26:     
cannam@26:     for (i=0;i<totalFilters;i++){
cannam@26:         triangleHeight[i] = 2./(upper[i]-lower[i]);
cannam@26:     }
cannam@26:   
cannam@26:     for (i=0;i<fftSize;i++){
cannam@26:         fftFreqs[i] = ((double) i / ((double) fftSize ) * 
cannam@26:                        (double) samplingRate);
cannam@26:     }
cannam@25: 
cannam@26:     /* Build now the mccFilterWeight matrix */
cannam@26:     for (i=0;i<totalFilters;i++){
cannam@25: 
cannam@26:         for (j=0;j<fftSize;j++) {
cannam@26:       
cannam@26:             if ((fftFreqs[j] > lower[i]) && (fftFreqs[j] <= center[i])) {
cannam@26:           
cannam@26:                 mfccFilterWeights[i][j] = triangleHeight[i] * 
cannam@26:                     (fftFreqs[j]-lower[i]) / (center[i]-lower[i]); 
cannam@26:           
cannam@26:             }
cannam@26:             else
cannam@26:             {
cannam@26:                 mfccFilterWeights[i][j] = 0.0;
cannam@26:             }
cannam@25: 
cannam@26:             if ((fftFreqs[j]>center[i]) && (fftFreqs[j]<upper[i])) {
cannam@25: 
cannam@26:                 mfccFilterWeights[i][j] = mfccFilterWeights[i][j] + triangleHeight[i] * (upper[i]-fftFreqs[j]) 
cannam@26:                     / (upper[i]-center[i]);
cannam@26:             }
cannam@26:             else
cannam@26:             {
cannam@26:                 mfccFilterWeights[i][j] = mfccFilterWeights[i][j] + 0.0;
cannam@26:             }
cannam@25:         }
cannam@25: 
cannam@26:     }
cannam@25: 
cannam@26:     /*
cannam@26:      * We calculate now mfccDCT matrix 
cannam@26:      * NB: +1 because of the DC component
cannam@26:      */
cannam@26:   
cannam@26:     for (i=0; i<nceps+1; i++) {
cannam@26:         for (j=0; j<totalFilters; j++) {
cannam@26:             mfccDCTMatrix[i][j] = (1./sqrt((double) totalFilters / 2.))  
cannam@26:                 * cos((double) i * ((double) j + 0.5) / (double) totalFilters * M_PI);
cannam@25:         }
cannam@25:     }
cannam@25: 
cannam@26:     for (j=0;j<totalFilters;j++){
cannam@26:         mfccDCTMatrix[0][j]     =  (sqrt(2.)/2.) * mfccDCTMatrix[0][j];
cannam@26:     }
cannam@26:    
cannam@26:     /* The analysis window */
cannam@26:     window = new Window<double>(HammingWindow, fftSize);
cannam@25: 
cannam@26:     /* Allocate memory for the FFT */
cannam@26:     imagIn      = (double*)calloc(fftSize,sizeof(double));
cannam@26:     realOut     = (double*)calloc(fftSize,sizeof(double));
cannam@26:     imagOut     = (double*)calloc(fftSize,sizeof(double));
cannam@25:   
cannam@26:     free(freqs);
cannam@26:     free(lower);
cannam@26:     free(center);
cannam@26:     free(upper);
cannam@26:     free(triangleHeight);
cannam@26:     free(fftFreqs);
cannam@25: }
cannam@25: 
cannam@26: MFCC::~MFCC()
cannam@26: {
cannam@26:     int i;
cannam@26:   
cannam@26:     /* Free the structure */
cannam@26:     for (i=0;i<nceps+1;i++) {
cannam@26:         free(mfccDCTMatrix[i]);
cannam@26:     }
cannam@26:     free(mfccDCTMatrix);
cannam@26:     
cannam@26:     for (i=0;i<totalFilters; i++) {
cannam@26:         free(mfccFilterWeights[i]);
cannam@26:     }
cannam@26:     free(mfccFilterWeights);
cannam@26:     
cannam@26:     /* Free the feature vector */
cannam@26:     free(ceps);
cannam@26:     
cannam@26:     /* The analysis window */
cannam@26:     delete window;
cannam@26:     
cannam@26:     /* Free the FFT */
cannam@26:     free(imagIn);
cannam@26:     free(realOut);
cannam@26:     free(imagOut);
cannam@26: }
cannam@25: 
cannam@25: 
cannam@25: /*
cannam@25:  * 
cannam@25:  * Extract the MFCC on the input frame 
cannam@26:  *
cannam@26:  * looks like we have to have length = fftSize ??????
cannam@25:  * 
cannam@25:  */ 
cannam@26: int MFCC::process(int length, double *inframe, double *outceps)
cannam@26: {
cannam@26:     int i,j;
cannam@26:   
cannam@26:     double *fftMag;
cannam@26:     double *earMag;
cannam@25: 
cannam@26:     double *inputData;
cannam@26:   
cannam@26:     double tmp;
cannam@25: 
cannam@26:     earMag    = (double*)calloc(totalFilters, sizeof(double));
cannam@26:     inputData = (double*)calloc(fftSize, sizeof(double)); 
cannam@26:     fftMag    = (double*)calloc(fftSize, sizeof(double)); 
cannam@26:     
cannam@26:     /* Zero-pad if needed */
cannam@26:     memcpy(inputData, inframe, length*sizeof(double));
cannam@25: 
cannam@26:     window->cut(inputData);
cannam@26:   
cannam@26:     /* Calculate the fft on the input frame */
cannam@26:     FFT::process(fftSize, 0, inputData, imagIn, realOut, imagOut);
cannam@25: 
cannam@26:     for (i = 0; i < fftSize; ++i) {
cannam@26:         fftMag[i] = sqrt(realOut[i] * realOut[i] +
cannam@26:                          imagOut[i] * imagOut[i]);
cannam@25:     }
cannam@25: 
cannam@26:     /* Multiply by mfccFilterWeights */
cannam@26:     for (i=0;i<totalFilters;i++) {
cannam@26:         tmp = 0.;
cannam@26:         for(j=0;j<fftSize/2; j++) {
cannam@26:             tmp = tmp + (mfccFilterWeights[i][j]*fftMag[j]);
cannam@26:         }
cannam@26:         if (tmp>0)
cannam@26:             earMag[i] = log10(tmp);
cannam@26: 	else
cannam@26:             earMag[i] = 0.0;
cannam@26:     }
cannam@26: 
cannam@26:     /*
cannam@26:      * 
cannam@26:      * Calculate now the cepstral coefficients 
cannam@26:      * with or without the DC component
cannam@26:      *
cannam@26:      */
cannam@26:   
cannam@26:     if (WANT_C0==1) {
cannam@26:      
cannam@26:         for (i=0;i<nceps+1;i++) {
cannam@26:             tmp = 0.;
cannam@26:             for (j=0;j<totalFilters;j++){
cannam@26:                 tmp = tmp + mfccDCTMatrix[i][j]*earMag[j];
cannam@26:             }
cannam@26:             outceps[i] = tmp;
cannam@26:         }
cannam@26:     }
cannam@25:     else 
cannam@26:     {  
cannam@26:         for (i=1;i<nceps+1;i++) {
cannam@26:             tmp = 0.;
cannam@26:             for (j=0;j<totalFilters;j++){
cannam@26:                 tmp = tmp + mfccDCTMatrix[i][j]*earMag[j];
cannam@26:             }
cannam@26:             outceps[i-1] = tmp;
cannam@25:         }
cannam@26:     }
cannam@25:     
cannam@26:     free(fftMag);
cannam@26:     free(earMag);
cannam@26:     free(inputData);
cannam@25: 
cannam@26:     return nceps;
cannam@25: }
cannam@25: