FormatClassifier.cpp

Go to the documentation of this file.

/**********************************************************************
 
  Audacity: A Digital Audio Editor
 
  FormatClassifier.cpp
 
  Philipp Sibler
 
******************************************************************//*******************************************************************/
#include "FormatClassifier.h"
 
#include <stdint.h>
#include <cmath>
#include <cfloat>
#include <vector>
#include <cstdio>
 
#include <wx/defs.h>
 
#include "sndfile.h"
 
FormatClassifier::FormatClassifier(const char* filename) :
   mReader(filename),
   mMeter(cSiglen)
{
   // Define the classification classes
   for ( auto endianness : {
      MachineEndianness::Little,
      MachineEndianness::Big,
   } )
      for ( auto format : {
         MultiFormatReader::Int8,
         MultiFormatReader::Int16,
         MultiFormatReader::Int32,
         MultiFormatReader::Uint8,
         MultiFormatReader::Float,
         MultiFormatReader::Double,
      } )
         mClasses.push_back( { format, endianness } );
 
   // Build feature vectors
   mMonoFeat = Floats{ mClasses.size() };
   mStereoFeat = Floats{ mClasses.size() };
   
#ifdef FORMATCLASSIFIER_SIGNAL_DEBUG
   // Build a debug writer
   char dfile [1024];
   sprintf(dfile, "%s.sig", filename);
   mpWriter = std::make_unique<DebugWriter>(dfile);
#endif
 
   // Run it
   Run();
   
#ifdef FORMATCLASSIFIER_SIGNAL_DEBUG
   for (unsigned int n = 0; n < mClasses.size(); n++)
   {
      wxPrintf("Class [%i] Machine [%i]: Mono: %3.7f Stereo: %3.7f\n", mClasses[n].format, mClasses[n].endian, mMonoFeat[n], mStereoFeat[n]);
   }
#endif
 
}
 
FormatClassifier::~FormatClassifier()
{
}
 
FormatClassifier::FormatClassT FormatClassifier::GetResultFormat()
{
   return mResultFormat;
}
 
int FormatClassifier::GetResultFormatLibSndfile()
{
   int format = SF_FORMAT_RAW;
   
   switch(mResultFormat.format)
   {
      case MultiFormatReader::Int8:
         format |= SF_FORMAT_PCM_S8;
         break;
      case MultiFormatReader::Int16:
         format |= SF_FORMAT_PCM_16;
         break;
      case MultiFormatReader::Int32:
         format |= SF_FORMAT_PCM_32;
         break;
      case MultiFormatReader::Uint8:
         format |= SF_FORMAT_PCM_U8;
         break;
      case MultiFormatReader::Float:
         format |= SF_FORMAT_FLOAT;
         break;
      case MultiFormatReader::Double:
         format |= SF_FORMAT_DOUBLE;
         break;
      default:
         format |= SF_FORMAT_PCM_16;
         break;
   }
   
   switch(mResultFormat.endian)
   {
      case MachineEndianness::Little:
         format |= SF_ENDIAN_LITTLE;
         break;
      case MachineEndianness::Big:
         format |= SF_ENDIAN_BIG;
         break;
   }
   
   return format;
}
 
unsigned FormatClassifier::GetResultChannels()
{
   return mResultChannels;
}
 
void FormatClassifier::Run()
{
   // Calc the mono feature vector
   for (unsigned int n = 0; n < mClasses.size(); n++)
   {
      // Read the signal
      ReadSignal(mClasses[n], 1);
#ifdef FORMATCLASSIFIER_SIGNAL_DEBUG
      mpWriter->WriteSignal(mSigBuffer, cSiglen);
#endif
 
      // Do some simple preprocessing
      // Remove DC offset
      float smean = Mean(mSigBuffer.get(), cSiglen);
      Sub(mSigBuffer.get(), smean, cSiglen);
      // Normalize to +- 1.0
      Abs(mSigBuffer.get(), mAuxBuffer.get(), cSiglen);
      float smax = Max(mAuxBuffer.get(), cSiglen);
      Div(mSigBuffer.get(), smax, cSiglen);
 
      // Now actually fill the feature vector
      // Low to high band power ratio
      float pLo = mMeter.CalcPower(mSigBuffer.get(), 0.15f, 0.3f);
      float pHi = mMeter.CalcPower(mSigBuffer.get(), 0.45f, 0.1f);
      mMonoFeat[n] = pLo / pHi;
   }
 
   // Calc the stereo feature vector
   for (unsigned int n = 0; n < mClasses.size(); n++)
   {
      // Read the signal
      ReadSignal(mClasses[n], 2);
#ifdef FORMATCLASSIFIER_SIGNAL_DEBUG
      mpWriter->WriteSignal(mSigBuffer, cSiglen);
#endif
 
      // Do some simple preprocessing
      // Remove DC offset
      float smean = Mean(mSigBuffer.get(), cSiglen);
      Sub(mSigBuffer.get(), smean, cSiglen);
      // Normalize to +- 1.0
      Abs(mSigBuffer.get(), mAuxBuffer.get(), cSiglen);
      float smax = Max(mAuxBuffer.get(), cSiglen);
      Div(mSigBuffer.get(), smax, cSiglen);
 
      // Now actually fill the feature vector
      // Low to high band power ratio
      float pLo = mMeter.CalcPower(mSigBuffer.get(), 0.15f, 0.3f);
      float pHi = mMeter.CalcPower(mSigBuffer.get(), 0.45f, 0.1f);
      mStereoFeat[n] = pLo / pHi;
   }
 
   // Get the results
   size_t midx, sidx;
   float monoMax = Max(mMonoFeat.get(), mClasses.size(), &midx);
   float stereoMax = Max(mStereoFeat.get(), mClasses.size(), &sidx);
 
   if (monoMax > stereoMax)
   {
      mResultChannels = 1;
      mResultFormat = mClasses[midx];
   }
   else
   {
      mResultChannels = 2;
      mResultFormat = mClasses[sidx];
   }
 
}
 
void FormatClassifier::ReadSignal(FormatClassT format, size_t stride)
{
   size_t actRead = 0;
   unsigned int n = 0;
 
   mReader.Reset();
 
   // Do a dummy read of 1024 bytes to skip potential header information
   mReader.ReadSamples(mRawBuffer.get(), 1024, MultiFormatReader::Uint8, MachineEndianness::Little);
 
   do
   {
      actRead = mReader.ReadSamples(mRawBuffer.get(), cSiglen, stride, format.format, format.endian);
 
      if (n == 0)
      {
         ConvertSamples(mRawBuffer.get(), mSigBuffer.get(), format);
      }
      else
      {
         if (actRead == cSiglen)
         {
            ConvertSamples(mRawBuffer.get(), mAuxBuffer.get(), format);
 
            // Integrate signals
            Add(mSigBuffer.get(), mAuxBuffer.get(), cSiglen);
 
            // Do some dummy reads to break signal coherence
            mReader.ReadSamples(mRawBuffer.get(), n + 1, stride, format.format, format.endian);
         }
      }
 
      n++;
 
   } while ((n < cNumInts) && (actRead == cSiglen));
 
}
 
void FormatClassifier::ConvertSamples(void* in, float* out, FormatClassT format)
{
   switch(format.format)
   {
      case MultiFormatReader::Int8:
         ToFloat((int8_t*) in, out, cSiglen);
         break;
      case MultiFormatReader::Int16:
         ToFloat((int16_t*) in, out, cSiglen);
         break;
      case MultiFormatReader::Int32:
         ToFloat((int32_t*) in, out, cSiglen);
         break;
      case MultiFormatReader::Uint8:
         ToFloat((uint8_t*) in, out, cSiglen);
         break;
      case MultiFormatReader::Uint16:
         ToFloat((uint16_t*) in, out, cSiglen);
         break;
      case MultiFormatReader::Uint32:
         ToFloat((uint32_t*) in, out, cSiglen);
        break;
      case MultiFormatReader::Float:
         ToFloat((float*) in, out, cSiglen);
         break;
      case MultiFormatReader::Double:
         ToFloat((double*) in, out, cSiglen);
         break;
   }
}
 
void FormatClassifier::Add(float* in1, float* in2, size_t len)
{
   for (unsigned int n = 0; n < len; n++)
   {
      in1[n] += in2[n];
   }
}
 
void FormatClassifier::Sub(float* in, float subt, size_t len)
{
   for (unsigned int n = 0; n < len; n++)
   {
      in[n] -= subt;
   }
}
 
void FormatClassifier::Div(float* in, float div, size_t len)
{
   for (unsigned int n = 0; n < len; n++)
   {
      in[n] /= div;
   }
}
 
 
void FormatClassifier::Abs(float* in, float* out, size_t len)
{
   for (unsigned int n = 0; n < len; n++)
   {
      if (in[n] < 0.0f)
      {
         out[n] = -in[n];
      }
      else
      {
         out[n] = in[n];
      }
   }
}
 
float FormatClassifier::Mean(float* in, size_t len)
{
   float mean = 0.0f;
 
   for (unsigned int n = 0; n < len; n++)
   {
      mean += in[n];
   }
 
   mean /= len;
   
   return mean;
}
 
float FormatClassifier::Max(float* in, size_t len)
{
   size_t dummyidx;
   return Max(in, len, &dummyidx);
}
 
float FormatClassifier::Max(float* in, size_t len, size_t* maxidx)
{
   float max = -FLT_MAX;
   *maxidx = 0;
   
   for (unsigned int n = 0; n < len; n++)
   {
      if (in[n] > max)
      {
         max = in[n];
         *maxidx = n;
      }
   }
 
   return max;
}
 
template<class T> void FormatClassifier::ToFloat(T* in, float* out, size_t len)
{
   for(unsigned int n = 0; n < len; n++)
   {
      out[n] = (float) in[n];
   }
}