FormantShifter.cpp

Go to the documentation of this file.

/*  SPDX-License-Identifier: GPL-2.0-or-later */
/*!********************************************************************
 
  Audacity: A Digital Audio Editor
 
  FormantShifter.cpp
 
  Matthieu Hodgkinson
 
**********************************************************************/
#include "FormantShifter.h"
#include "FormantShifterLoggerInterface.h"
#include "MapToPositiveHalfIndex.h"
#include "MathApprox.h"
#include <algorithm>
#include <cassert>
#include <fstream>
 
namespace
{
// `x` has length `fftSize/2+1`.
// Returns the last bin that wasn't zeroed.
size_t ResampleFreqDomain(float* x, size_t fftSize, double factor)
{
   const auto size = fftSize / 2 + 1;
   const auto end = std::min(size, size_t(size * factor));
   std::vector<float> tmp(end);
   for (size_t i = 0; i < end; ++i)
   {
      const int int_pos = i / factor;
      const float frac_pos = 1.f * i / factor - int_pos;
      const auto k = MapToPositiveHalfIndex(int_pos, fftSize);
      const auto l = MapToPositiveHalfIndex(int_pos + 1, fftSize);
      tmp[i] = (1 - frac_pos) * x[k] + frac_pos * x[l];
   }
   std::copy(tmp.begin(), tmp.begin() + end, x);
   if (end < size)
      std::fill(x + end, x + size, 0.f);
   return end;
}
} // namespace
 
FormantShifter::FormantShifter(
   int sampleRate, double cutoffQuefrency,
   FormantShifterLoggerInterface& logger)
    : cutoffQuefrency { cutoffQuefrency }
    , mSampleRate { sampleRate }
    , mLogger { logger }
{
}
 
void FormantShifter::Reset(size_t fftSize)
{
   mFft = std::make_unique<staffpad::audio::FourierTransform>(
      static_cast<int32_t>(fftSize));
   const auto numBins = fftSize / 2 + 1;
   mEnvelope.setSize(1, numBins);
   mCepstrum.setSize(1, fftSize);
   mEnvelopeReal.resize(numBins);
   mWeights.resize(numBins);
}
 
void FormantShifter::Reset()
{
   mFft.reset();
}
 
void FormantShifter::Process(
   const float* powSpec, std::complex<float>* spec, double factor)
{
   assert(factor > 0);
   if (factor <= 0 || cutoffQuefrency == 0 || !mFft)
      return;
 
   const auto fftSize = mFft->getSize();
   const auto numBins = fftSize / 2 + 1;
 
   mLogger.Log(fftSize, "fftSize");
 
   // Take the log of the normalized magnitude. (This assumes that
   // the window averages to 1.)
   std::complex<float>* pEnv = mEnvelope.getPtr(0);
   const float normalizer = FastLog2(fftSize);
   std::transform(powSpec, powSpec + numBins, pEnv, [&](float power) {
      return .5f * FastLog2(power) - normalizer;
   });
 
   // Get the cosine transform of the log magnitude, aka the cepstrum.
   mFft->inverseReal(mEnvelope, mCepstrum);
   auto pCepst = mCepstrum.getPtr(0);
   mLogger.Log(pCepst, fftSize, "cepstrum");
 
   // "Lifter" the cepstrum.
   const auto binCutoff = int(mSampleRate * cutoffQuefrency * factor);
   if (binCutoff < fftSize / 2)
      std::fill(pCepst + binCutoff + 1, pCepst + fftSize - binCutoff, 0.f);
   mLogger.Log(pCepst, fftSize, "cepstrumLiftered");
 
   // Get the envelope back.
   mFft->forwardReal(mCepstrum, mEnvelope);
   std::transform(
      pEnv, pEnv + numBins, mEnvelopeReal.begin(),
      [fftSize = fftSize](const std::complex<float>& env) {
         return std::exp2(env.real() / fftSize);
      });
   mLogger.Log(mEnvelopeReal.data(), numBins, "envelope");
 
   // Get the weights, which are the ratio of the desired envelope to the
   // current envelope (which has the effect of downsampling).
   std::transform(
      mEnvelopeReal.begin(), mEnvelopeReal.end(), mWeights.begin(),
      [](float env) { return std::isnormal(env) ? 1.f / env : 0.f; });
 
   const auto lastNonZeroedBin =
      ResampleFreqDomain(mEnvelopeReal.data(), fftSize, factor);
 
   mLogger.Log(mEnvelopeReal.data(), numBins, "envelopeResampled");
   std::transform(
      mEnvelopeReal.begin(), mEnvelopeReal.end(), mWeights.begin(),
      mWeights.begin(), [](float env, float weight) {
         // Limit the weights to 100, which corresponds to 20dB.
         // Our purpose is to add (or remove) energy to formants, and it doesn't
         // need to be by more than that. This way we also avoid unreasonable
         // amplification.
         return std::min(env * weight, 100.f);
      });
 
   // Say the signal was downsampled to pitch it up. The factor is then less
   // than 1, and the resampler had to zero out the upper part of the envelope
   // bins. For these, rather than zeroing the spec too, it sounds better
   // to keep the original, even if no envelope correction is applied, else the
   // signal looses a bit of clarity. At such high frequencies, we probably
   // don't need a smooth frequency-domain transition and a jump is fine. (This
   // is visible in the spec, in case you're curious.)
   std::fill(mWeights.begin() + lastNonZeroedBin, mWeights.end(), 1.f);
 
   mLogger.Log(mWeights.data(), mWeights.size(), "weights");
 
   mLogger.Log(
      spec, numBins, "magnitude",
      [fftSize = fftSize](const std::complex<float>& spec) {
         return std::abs(spec) / fftSize;
      });
 
   // Now apply the weights.
   std::transform(
      spec, spec + numBins, mWeights.begin(), spec,
      std::multiplies<std::complex<float>>());
 
   mLogger.Log(
      spec, numBins, "weightedMagnitude",
      [fftSize = fftSize](const std::complex<float>& spec) {
         return std::abs(spec) / fftSize;
      });
 
   mLogger.ProcessFinished(spec, fftSize);
}