Crash in Dio

[3628800]

Hi, there is a crash calling Dio (manifests especially for short signals) when the low-cut filter frequency is large compared to the calculated DFT size.

Specifically, the crash occurs in GetSpectrumForEstimation when fft_size < (cutoff_in_sample * 2 + 1), and DesignLowCutFilter(cutoff_in_sample * 2 + 1, fft_size, y) indexes y (created with length fft_size) upto the first parameter.

[mmorise]

Thank you for your suggestion.

I think that the following modification can solve the problem. We cannot use the low-cut filter in this case, but I don't have another idea to simply solve it. Perhaps, this error rarely occurs, so its simple approach would be reasonable. If you have another idea, please propose it as a pull request. I will consider your idea.

if (cutoff_in_sample * 2 + 1 <= fft_size) DesignLowCutFilter(cutoff_in_sample * 2 + 1, fft_size, y);

[3628800]

Thanks for the fast response.

A modification to your modification: I think the test needs to be cutoff_in_sample * 3 + 1 <= fft_size based on the maximum loop index in DesignLowCutFilter of (N-1) + (N-1)/2 (in the last loop).

Must the filtering operation be skipped in this case, or does it make any sense to filter the signal up to min(cutoff_in_sample * 2 + 1, 2/3. * (.5 + fft_size))?

[mmorise]

Thank you very much for your comment. Sorry, I'm too busy to modify the program until this weekend. I will modify the program next week.

[3628800]

Thanks for your time!

[mmorise]

I checked the source code. Based on the result, I would like to propose another idea. In line 592 of dio.cpp, fft_size is determined by the following equation. int fft_size = GetSuitableFFTSize(y_length + (4 * static_cast<int>(1.0 + actual_fs / boundary_f0_list[0] / 2.0))); This equation does not include the length of the low-cut filter. So, the following modification would be effective. int fft_size = GetSuitableFFTSize(y_length + matlab_round(actual_fs / 50.0) * 2 + 1 + (4 * static_cast<int>(1.0 + actual_fs / boundary_f0_list[0] / 2.0))); This equation guarantees that the fft_size is always longer than the length of low-cut filter.

Since the value 50.0 is constant, the modification should also include the additional definition in the constantnumbers.h. const double kLowCut = 50.0; int fft_size = GetSuitableFFTSize(y_length + matlab_round(actual_fs / world::kLowCut) * 2 + 1 + (4 * static_cast<int>(1.0 + actual_fs / boundary_f0_list[0] / 2.0)));

If you have no comment, I will modify the source codes (dio.cpp and constantnumbers.h).

[3628800]

That seems fine to me. (Although for my particular case this results in an 8x increase in the FFT size, so it might be worth testing on some of your "normal use-case" inputs to make sure the library won't suffer a dramatic slowdown...!)

[mmorise]

I update the source codes in dio.cpp and constantnumbers.h.

In cases where the length of input signal is long, the processing speed would be little affected.

[3628800]

Thank you.