the realtime' result is different from the no-realtime result.

[bringtree]

Hello, I try to align the realtime result and no-realtime result.

I have been replaced the "rand()" with the const value. And I find the realtime's value（by WaveformSynthesis2 or WaveformSynthesis3） is different from the no-realtime value (by WaveformSynthesis).

[bringtree]

For example, in synthesis.cpp

  double coefficient = 2.0 * world::kPi * fractional_time_shift * fs / fft_size;
  GetSpectrumWithFractionalTimeShift(fft_size, coefficient, inverse_real_fft);

btw, in synthesisrealtime.cpp, I can't find this process.

[mmorise]

Thank you for your comment.

There are several different points between real-time and offline synthesizers. One is the fractional component, as you pointed out, and another is the DC removal. The last point is the calculation of the temporal positions of the vocal cord vibrations. Since there is a rounding off, a temporal position is often different from an offline synthesizer.

[bringtree]

@mmorise thx, I also find several different points.

1. fft_size? fft_size/2? In synthesis.cpp

   GetDCRemover(fft_size, dc_remover);

In synthesisreadltime.cpp

 GetDCRemover(synth->fft_size / 2, synth->dc_remover);

In synthesis.cpp

static void RemoveDCComponent(const double *periodic_response, int fft_size,
                              const double *dc_remover,
                              double *new_periodic_response) {
  double dc_component = 0.0;
  for (int i = fft_size / 2; i < fft_size; ++i)
    dc_component += periodic_response[i];
  for (int i = 0; i < fft_size / 2; ++i)
    new_periodic_response[i] = -dc_component * dc_remover[i];
  for (int i = fft_size / 2; i < fft_size; ++i)
    new_periodic_response[i] -= dc_component * dc_remover[i];
}

In synthesisreadltime.cpp

[mmorise]

Yes, DC removal of real-time synthesis is different from offline synthesis because of the causality. On the other hand, the influence of the sound quality is negligible.

[bringtree]

@mmorise thx, and about buffer size(64). Is the buffer size related to the sampling rate(24k)？ I found the "buffer size" makes a difference in wav result.

[mmorise]

Is the buffer size related to the sampling rate(24k)？

No, you can set the buffer size independently. I think that the difference may be caused by the rounding off in the struct. The calculation of the temporal position of vocal cord vibration may change due to this difference.