Changeset e5f6a0b

Timestamp:

Sep 16, 2007, 6:06:07 PM (17 years ago)

Author:

Paul Brossier <piem@piem.org>

Branches:

feature/autosink, feature/cnn, feature/cnn_org, feature/constantq, feature/crepe, feature/crepe_org, feature/pitchshift, feature/pydocstrings, feature/timestretch, fix/ffmpeg5, master, pitchshift, sampler, timestretch, yinfft+

Children:

177f09a

Parents:

45134c5 (diff), bc4ba75 (diff)
Note: this is a merge changeset, the changes displayed below correspond to the merge itself.
Use the (diff) links above to see all the changes relative to each parent.

Message:

merge from amaury's branch

Files:

• examples/aubiomfcc.c (modified) (4 diffs)
• examples/plotfb.py (added)
• examples/plotmat.py (added)
• examples/utils.c (modified) (5 diffs)
• src/filterbank.c (modified) (4 diffs)
• src/filterbank.h (modified) (2 diffs)
• src/mfcc.c (modified) (4 diffs)
• src/mfcc.h (modified) (2 diffs)
• src/sample.c (modified) (1 diff)
• swig/aubio.i (modified) (1 diff)

examples/aubiomfcc.c

@@ -23 +23 @@
 fvec_t * mfcc_out;
 aubio_mfcc_t * mfcc;
+
+uint_t n_filters = 40;
+uint_t n_coefs = 11;
 
 unsigned int pos = 0; /*frames%dspblocksize*/
@@ -50 +53 @@
       //compute mfccs
       aubio_mfcc_do(mfcc, fftgrain, mfcc_out);
-
+      
+      uint_t coef_cnt;
+      for (coef_cnt = 0; coef_cnt < n_coefs; coef_cnt++) {
+          outmsg("%f ",mfcc_out->data[0][coef_cnt]);
+      }
+      outmsg("\n");
+      
       /* end of block loop */
       pos = -1; /* so it will be zero next j loop */
@@ -65 +74 @@
       */
       
-      uint_t filter_cnt;
+      uint_t coef_cnt;
       if (output_filename == NULL) {
-        if(frames >= 4) {
-          outmsg("%f\t",(frames-4)*overlap_size/(float)samplerate);
-        } else if (frames < 4) {
-          outmsg("%f\t",0.);
-        }
-        outmsg("%f",mfcc_out->data[0][0]);
-        for (filter_cnt = 1; filter_cnt < mfcc_out->length; filter_cnt++) {
-          outmsg(",%f",mfcc_out->data[0][filter_cnt]);
-        }
-        outmsg("\n");
+//         if(frames >= 4) {
+//           outmsg("%f\t",(frames-4)*overlap_size/(float)samplerate);
+//         } 
+//         else if (frames < 4) {
+//           outmsg("%f\t",0.);
+//         }
+        //outmsg("%f ",mfcc_out->data[0][0]);
+        
+        
       }
 }
@@ -82 +90 @@
 int main(int argc, char **argv) {
   // params
-  uint_t n_filters = 11;
+  
   examples_common_init(argc,argv);
-  smpl_t lowfreq = 0.;
-  smpl_t highfreq = samplerate;
-  mfcc_out = new_fvec(n_filters,channels);
+  smpl_t lowfreq = 133.333f;
+  smpl_t highfreq = 44100.f;
+  mfcc_out = new_fvec(n_coefs,channels);
+  
   
   //populating the filter
-  mfcc = new_aubio_mfcc(buffer_size, samplerate, n_filters, lowfreq, highfreq,
-      channels);
-
+  mfcc = new_aubio_mfcc(buffer_size, samplerate, n_filters, n_coefs , lowfreq, highfreq, channels);
+  dump_filterbank(mfcc);
+  
+  
   //process
   examples_common_process(aubio_process,process_print);

examples/utils.c

@@ -40 +40 @@
 smpl_t threshold                      = 0.3;
 smpl_t silence                        = -90.;
-uint_t buffer_size                    = 512; //1024;
-uint_t overlap_size                   = 256; //512;
+// uint_t buffer_size                    = 512;
+// uint_t overlap_size                   = 256;
+uint_t buffer_size                    = 1024;
+uint_t overlap_size                   = 512;
+// uint_t buffer_size                    = 2048;
+// uint_t overlap_size                   = 1024;
 uint_t channels                       = 1;
 uint_t samplerate                     = 44100;
@@ -60 +64 @@
 int isonset = 0;
 aubio_pickpeak_t * parms;
-
-/* mfcc objects */
-//parameters
-uint_t n_filters=20;
-smpl_t lowfreq=80.f;
-smpl_t highfreq=18000.f;
-// filterbank object
-aubio_filterbank_t * mf;
-
-// DCT mfft and result storage
-aubio_mfft_t * fft_dct;
-cvec_t * fftgrain_dct;
-smpl_t * mfcc_outbuf[11];
-
 
 /* pitch objects */
@@ -314 +304 @@
   fftgrain  = new_cvec(buffer_size, channels);
 
-  //init for mfcc process
-  fftgrain_dct= new_cvec(n_filters, channels);
-
+  
   if (usepitch) {
     pitchdet = new_aubio_pitchdetection(buffer_size*4, 
@@ -330 +318 @@
   pv = new_aubio_pvoc(buffer_size, overlap_size, channels);
   
-  // dct phase vocoder
-  //TODO: check size
-  fft_dct = new_aubio_mfft(n_filters, channels);
-
   /* onsets */
   parms = new_aubio_peakpicker(threshold);
@@ -367 +351 @@
   del_fvec(onset);
   del_fvec(woodblock);
-  
-  //mffc related
-  del_aubio_mfft(fft_dct);
-  del_cvec(fftgrain_dct);
   
   aubio_cleanup();

src/filterbank.c

@@ -19 +19 @@
 */
 
-/* part of this mfcc implementation were inspired from LibXtract
-   http://libxtract.sourceforge.net/
-*/
 
 #include "aubio_priv.h"
 #include "sample.h"
 #include "filterbank.h"
+
+#include "stdio.h"
 
 #define USE_EQUAL_GAIN 1
@@ -53 +52 @@
 }
 
-aubio_filterbank_t * new_aubio_filterbank_mfcc(uint_t n_filters, uint_t win_s, smpl_t samplerate, smpl_t freq_min, smpl_t freq_max){
+aubio_filterbank_t * new_aubio_filterbank_mfcc(uint_t n_filters, uint_t win_s, uint_t samplerate, smpl_t freq_min, smpl_t freq_max){
+  
   smpl_t nyquist = samplerate/2.;
   uint_t style = 1;
@@ -142 +142 @@
     for(k = next_peak + 1; k < fb->win_s; k++)
       fb->filters[n]->data[0][k]=0.f;
+
+
   }
 
@@ -149 +151 @@
   free(fft_peak);
 
+
   return fb;
 
 }
 
+/*
+FB initialization based on Slaney's auditory toolbox
+TODO:
+  *solve memory leak problems while
+  *solve quantization issues when constructing signal:
+    *bug for win_s=512
+    *corrections for win_s=1024 -> why even filters with smaller amplitude
+
+*/
+
+aubio_filterbank_t * new_aubio_filterbank_mfcc2(uint_t n_filters, uint_t win_s, uint_t samplerate, smpl_t freq_min, smpl_t freq_max){
+  
+  aubio_filterbank_t * fb = new_aubio_filterbank(n_filters, win_s);
+  
+  
+  //slaney params
+  smpl_t lowestFrequency = 133.3333;
+  smpl_t linearSpacing = 66.66666666;
+  smpl_t logSpacing = 1.0711703;
+
+  uint_t linearFilters = 13;
+  uint_t logFilters = 27;
+  uint_t allFilters = linearFilters + logFilters;
+  
+  //buffers for computing filter frequencies
+  fvec_t * freqs=new_fvec(allFilters+2 , 1);
+  
+  fvec_t * lower_freqs=new_fvec( allFilters, 1);
+  fvec_t * upper_freqs=new_fvec( allFilters, 1);
+  fvec_t * center_freqs=new_fvec( allFilters, 1);
+
+  
+  fvec_t * triangle_heights=new_fvec( allFilters, 1);
+  //lookup table of each bin frequency in hz
+  fvec_t * fft_freqs=new_fvec(win_s, 1);
+
+  uint_t filter_cnt, bin_cnt;
+  
+  //first step: filling all the linear filter frequencies
+  for(filter_cnt=0; filter_cnt<linearFilters; filter_cnt++){
+    freqs->data[0][filter_cnt]=lowestFrequency+ filter_cnt*linearSpacing;
+  }
+  smpl_t lastlinearCF=freqs->data[0][filter_cnt-1];
+  
+  //second step: filling all the log filter frequencies
+  for(filter_cnt=0; filter_cnt<logFilters+2; filter_cnt++){
+    freqs->data[0][filter_cnt+linearFilters]=lastlinearCF*(pow(logSpacing,filter_cnt+1));
+  }
+
+  //Option 1. copying interesting values to lower_freqs, center_freqs and upper freqs arrays
+  //TODO: would be nicer to have a reference to freqs->data, anyway we do not care in this init step
+    
+  for(filter_cnt=0; filter_cnt<allFilters; filter_cnt++){
+    lower_freqs->data[0][filter_cnt]=freqs->data[0][filter_cnt];
+    center_freqs->data[0][filter_cnt]=freqs->data[0][filter_cnt+1];
+    upper_freqs->data[0][filter_cnt]=freqs->data[0][filter_cnt+2];
+  }
+
+
+  //computing triangle heights so that each triangle has unit area
+  for(filter_cnt=0; filter_cnt<allFilters; filter_cnt++){
+    triangle_heights->data[0][filter_cnt]=2./(upper_freqs->data[0][filter_cnt]-lower_freqs->data[0][filter_cnt]);
+  }
+  
+  
+  //AUBIO_DBG
+  AUBIO_DBG("filter tables frequencies\n");
+  for(filter_cnt=0; filter_cnt<allFilters; filter_cnt++)
+    AUBIO_DBG("filter n. %d %f %f %f %f\n",filter_cnt, lower_freqs->data[0][filter_cnt], center_freqs->data[0][filter_cnt], upper_freqs->data[0][filter_cnt], triangle_heights->data[0][filter_cnt]);
+
+  //filling the fft_freqs lookup table, which assigns the frequency in hz to each bin
+  for(bin_cnt=0; bin_cnt<win_s; bin_cnt++){
+    //TODO: check the formula!
+    fft_freqs->data[0][bin_cnt]= (smpl_t)samplerate* (smpl_t)bin_cnt/ (smpl_t)win_s;
+  }
+
+  //building each filter table
+  for(filter_cnt=0; filter_cnt<allFilters; filter_cnt++){
+
+    //TODO:check special case : lower freq =0
+    //calculating rise increment in mag/Hz
+    smpl_t riseInc= triangle_heights->data[0][filter_cnt]/(center_freqs->data[0][filter_cnt]-lower_freqs->data[0][filter_cnt]);
+    
+
+    AUBIO_DBG("\nfilter %d",filter_cnt);
+    //zeroing begining of filter
+    AUBIO_DBG("\nzero begin\n");
+    for(bin_cnt=0; bin_cnt<win_s-1; bin_cnt++){
+      //zeroing beigining of array
+      fb->filters[filter_cnt]->data[0][bin_cnt]=0.f;
+      AUBIO_DBG(".");
+      //AUBIO_DBG("%f %f %f\n", fft_freqs->data[0][bin_cnt], fft_freqs->data[0][bin_cnt+1], lower_freqs->data[0][filter_cnt]); 
+      if(fft_freqs->data[0][bin_cnt]<= lower_freqs->data[0][filter_cnt] && fft_freqs->data[0][bin_cnt+1]> lower_freqs->data[0][filter_cnt]){
+        break;
+      }
+    }
+    bin_cnt++;
+    
+    AUBIO_DBG("\npos slope\n");
+    //positive slope
+    for(; bin_cnt<win_s-1; bin_cnt++){
+      AUBIO_DBG(".");
+      fb->filters[filter_cnt]->data[0][bin_cnt]=(fft_freqs->data[0][bin_cnt]-lower_freqs->data[0][filter_cnt])*riseInc;
+      //if(fft_freqs->data[0][bin_cnt]<= center_freqs->data[0][filter_cnt] && fft_freqs->data[0][bin_cnt+1]> center_freqs->data[0][filter_cnt])
+      if(fft_freqs->data[0][bin_cnt+1]> center_freqs->data[0][filter_cnt])
+        break;
+    }
+    //bin_cnt++;
+    
+    
+    //negative slope
+    AUBIO_DBG("\nneg slope\n");   
+    for(; bin_cnt<win_s-1; bin_cnt++){
+      //AUBIO_DBG(".");
+      
+      //checking whether last value is less than 0...
+      smpl_t val=triangle_heights->data[0][filter_cnt]-(fft_freqs->data[0][bin_cnt]-center_freqs->data[0][filter_cnt])*riseInc;
+      if(val>=0)
+        fb->filters[filter_cnt]->data[0][bin_cnt]=val;
+      else fb->filters[filter_cnt]->data[0][bin_cnt]=0.f;
+      
+      //if(fft_freqs->data[0][bin_cnt]<= upper_freqs->data[0][bin_cnt] && fft_freqs->data[0][bin_cnt+1]> upper_freqs->data[0][filter_cnt])
+      //TODO: CHECK whether bugfix correct
+      if(fft_freqs->data[0][bin_cnt+1]> upper_freqs->data[0][filter_cnt])
+        break;
+    }
+    //bin_cnt++;
+    
+    AUBIO_DBG("\nzero end\n");
+    //zeroing tail
+    for(; bin_cnt<win_s; bin_cnt++)
+      //AUBIO_DBG(".");
+      fb->filters[filter_cnt]->data[0][bin_cnt]=0.f;
+
+  }
+  
+  
+
+  del_fvec(freqs);
+  del_fvec(lower_freqs);
+  del_fvec(upper_freqs);
+  del_fvec(center_freqs);
+
+  del_fvec(triangle_heights);
+  del_fvec(fft_freqs);
+
+  return fb;
+
+}
+
+void aubio_dump_filterbank(aubio_filterbank_t * fb){
+
+  FILE * mlog;
+  mlog=fopen("filterbank.txt","w");
+  
+  int k,n;
+  //dumping filter values
+  //smpl_t area_tmp=0.f;
+  for(n = 0; n < fb->n_filters; n++){
+    for(k = 0; k < fb->win_s; k++){
+      fprintf(mlog,"%f ",fb->filters[n]->data[0][k]);
+    }
+    fprintf(mlog,"\n");
+  }
+  
+  if(mlog) fclose(mlog);
+}
 
 void del_aubio_filterbank(aubio_filterbank_t * fb){

src/filterbank.h

@@ -46 +46 @@
 /** filterbank initialization for mel filters
 
-  \param nyquist nyquist frequency, i.e. half of the sampling rate
-  \param style libxtract style
-  \param freqmin lowest filter frequency
-  \param freqmax highest filter frequency
+  
+  \param n_filters number of filters
+  \param win_s window size
+  \param samplerate
+  \param freq_min lowest filter frequency
+  \param freq_max highest filter frequency
 
 */
-aubio_filterbank_t * new_aubio_filterbank_mfcc(uint_t n_filters, uint_t win_s, smpl_t samplerate, smpl_t freq_min, smpl_t freq_max);
+aubio_filterbank_t * new_aubio_filterbank_mfcc(uint_t n_filters, uint_t win_s, uint_t samplerate, smpl_t freq_min, smpl_t freq_max);
+
+/** filterbank initialization for mel filters
+
+  \param n_filters number of filters
+  \param win_s window size
+  \param samplerate
+  \param freq_min lowest filter frequency
+  \param freq_max highest filter frequency
+
+*/
+aubio_filterbank_t * new_aubio_filterbank_mfcc_2(uint_t n_filters, uint_t win_s, uint_t samplerate, smpl_t freq_min, smpl_t freq_max);
 
 
@@ -67 +80 @@
 void aubio_filterbank_do(aubio_filterbank_t * fb, cvec_t * in, fvec_t *out);
 
+/** dump filterbank filter tables in a txt file
+
+*/
+void aubio_dump_filterbank(aubio_filterbank_t * fb);
+
 #ifdef __cplusplus
 }

src/mfcc.c

@@ -34 +34 @@
   uint_t samplerate;        /** sample rate (needed?) */
   uint_t channels;          /** number of channels */
-  uint_t n_coefs;           /** number of coefficients (= fb->n_filters/2 +1) */
+  uint_t n_filters;         /** number of  *filters */
+  uint_t n_coefs;           /** number of coefficients (<= n_filters/2 +1) */
   smpl_t lowfreq;           /** lowest frequency for filters */ 
   smpl_t highfreq;          /** highest frequency for filters */
@@ -44 +45 @@
 
 
-aubio_mfcc_t * new_aubio_mfcc (uint_t win_s, uint_t samplerate ,uint_t n_coefs, smpl_t lowfreq, smpl_t highfreq, uint_t channels){
+aubio_mfcc_t * new_aubio_mfcc (uint_t win_s, uint_t samplerate, uint_t n_filters, uint_t n_coefs, smpl_t lowfreq, smpl_t highfreq, uint_t channels){
   /** allocating space for mfcc object */
   aubio_mfcc_t * mfcc = AUBIO_NEW(aubio_mfcc_t);
 
   //we need (n_coefs-1)*2 filters to obtain n_coefs coefficients after dct
-  uint_t n_filters = (n_coefs-1)*2;
+  //uint_t n_filters = (n_coefs-1)*2;
   
   mfcc->win_s=win_s;
   mfcc->samplerate=samplerate;
   mfcc->channels=channels;
+  mfcc->n_filters=n_filters;
   mfcc->n_coefs=n_coefs;
   mfcc->lowfreq=lowfreq;
   mfcc->highfreq=highfreq;
 
+  
   /** filterbank allocation */
-  mfcc->fb = new_aubio_filterbank_mfcc(n_filters, mfcc->win_s, samplerate, lowfreq, highfreq);
+  mfcc->fb = new_aubio_filterbank_mfcc2(n_filters, mfcc->win_s, samplerate, lowfreq, highfreq);
 
   /** allocating space for fft object (used for dct) */
@@ -102 +105 @@
     aubio_mfft_do (mf->fft_dct, in, mf->fftgrain_dct);
     //extract real part of fft grain
-    //for(i=0; i<mf->n_coefs ;i++){
-    for(i=0; i<out->length;i++){
+    for(i=0; i<mf->n_coefs ;i++){
+    //for(i=0; i<out->length;i++){
       out->data[0][i]= mf->fftgrain_dct->norm[0][i]
         *COS(mf->fftgrain_dct->phas[0][i]);
@@ -110 +113 @@
 }
 
+//a bit a kludge
+void dump_filterbank(aubio_mfcc_t * mf){
+  aubio_dump_filterbank(mf->fb);
+
+}
+

src/mfcc.h

@@ -45 +45 @@
 
 */
-aubio_mfcc_t * new_aubio_mfcc (uint_t win_s, uint_t samplerate ,uint_t n_coefs, smpl_t lowfreq, smpl_t highfreq, uint_t channels);
+aubio_mfcc_t * new_aubio_mfcc (uint_t win_s, uint_t samplerate, uint_t n_filters, uint_t n_coefs, smpl_t lowfreq, smpl_t highfreq, uint_t channels);
 /** delete mfcc object
 
@@ -70 +70 @@
 void aubio_dct_do(aubio_mfcc_t * mf, fvec_t *in, fvec_t *out);
 
+/** dump filterbank to log file
+
+  \param mf mfcc object as returned by new_aubio_mfcc
+  
+*/
+void dump_filterbank(aubio_mfcc_t * mf);
+
 #ifdef __cplusplus
 }

src/sample.c

@@ -20 +20 @@
 #include "aubio_priv.h"
 #include "sample.h"
+
 
 fvec_t * new_fvec( uint_t length, uint_t channels) {

swig/aubio.i

@@ -148 +148 @@
 smpl_t aubio_zero_crossing_rate(fvec_t * input);
 smpl_t aubio_spectral_centroid(cvec_t * spectrum, smpl_t samplerate);
+
+/* filterbank */
+
+/* mfcc */
+aubio_mfcc_t * new_aubio_mfcc (uint_t win_s, uint_t samplerate, uint_t n_filters, uint_t n_coefs, smpl_t lowfreq, smpl_t highfreq, uint_t channels);
+void del_aubio_mfcc(aubio_mfcc_t *mf);
+void aubio_mfcc_do(aubio_mfcc_t *mf, cvec_t *in, fvec_t *out);
+
 
 /* scale */