I thought I was nearly there but can't quite get it over the line.

[DrJaymz]

My goal was to work out the loudest amplitude fundamental from my samples and return its frequency. Using esp32.

Obviously before I start I got my I2S device recording and playing back - I can hear myself so thats working.

Then I needed to convert my data to double. Presumably so that it has positive and negative around 0 values.
I checked this for sanity and it also looks correct. i.e. you get a waveform centred on 0.

So then, it should just be a matter of run the FFT over my samples and use the FindMajorPeak but it fails on the FFT compute step.

I have 1024 samples at this point, block size = 1, sample rate is 22000 but should be enough to run compute and get something sensible.

TBH, I know what FFT does, I'm not great at working out how big the samples needed to be and I'm used to higher level languages where I don't need to be careful with the allocation and I don't need to get in to pointers; so these are probably where I'm getting it wrong if that helps?

data_offset = 0;
  InitI2SSpeakOrMic(MODE_MIC);
  size_t byte_read;
  while (1)
  {
    i2s_read(Speak_I2S_NUMBER, (char *)(microphonedata0 + data_offset),
             DATA_SIZE, &byte_read, (100 / portTICK_RATE_MS));
    data_offset += DATA_SIZE;
    if (data_offset == DATA_SIZE * BLOCKS)
      break;
  }

  // Can I play this back - Yes this plays it back ok.
  
  // size_t bytes_written;
  // InitI2SSpeakOrMic(MODE_SPK);
  // i2s_write(Speak_I2S_NUMBER, microphonedata0, data_offset, &bytes_written,
  //           portMAX_DELAY);

  // // Convert 8-bit unsigned int data to double for some reason

  double *micData = (double *)malloc((DATA_SIZE * BLOCKS) * sizeof(double));
  double *micDataImag = (double *)malloc((DATA_SIZE * BLOCKS) * sizeof(double));
  for (int i = 0; i < DATA_SIZE * BLOCKS; i++)
  {
    micData[i] = (double)(microphonedata0[i] - 128);
    micDataImag = 0;
  }

  // Apply Hamming window to microphone data
   FFT.Windowing(FFT_WIN_TYP_HAMMING, FFT_FORWARD);  // <-- crashes here LoadProhibited
   FFT.Compute(micData, micDataImag, DATA_SIZE * BLOCKS, FFT_FORWARD);

   double peak = FFT.MajorPeak(micData, DATA_SIZE * BLOCKS, SAMPLE_RATE);
   Serial.printf("Peak: %0.1f \n", peak);

   free(micData);
   free(micDataImag);

[DrJaymz]

micDataImag = 0; <--- duh.
micDataImag[i] = 0;

[blaz-r]

Hello,
so with this fix, the code is working as it should now?

Also, if you are doing this in a loop, it might be worth initializing memory outside of it, so you don't allocate and free every iteration, as that is just wasting time. Also, you can use new in c++ and then delete[] to deallocate it, which is a bit nicer than malloc. Or you can init array on stack/data segment without malloc or new if you know the size beforehand, like: double data[size];

[DrJaymz]

Well not really with that code. I found that it worked if I specified on the constructor as below but not as I showed it there;

double vReal[512];
double vImag[512];
arduinoFFT FFT = arduinoFFT(vReal, vImag, 512, 64000);

Then call:

  FFT.Compute(FFT_FORWARD);
  FFT.ComplexToMagnitude();
  double peak = FFT.MajorPeakParabola(); 

Then I get something sensible. I wasn't sure what FFT functions I needed to call before FFT.MajorPeakParabola();
But this seemed to work. No idea if I needed to remove DC or not beforehand, but I don't think it changes anything.

There were other problems with the I2S sampling. I just wanted to get to the point where the mic was being sampled via I2s continually and I could just drop in when I felt like it and grab 512 samples and get the most dominant frequency.
I have it working but its a bit messy to integrate with my existing code right now.

What I really need is a bandpass filter, which I think I will can do on the samples directly before passing into fft.

[kosme]

@DrJaymz, I see a problem with the setup you describe. The sampling rate should be at least twice the highest expected frequency. Since you mention that you can hear yourself speaking, I infer you are dealing with audio. Therefore, the sampling frequency should be at least 44000.

[DrJaymz]

The sampling frequency is 64000hz. I am using the m5stack core 2 built in spm1423 microphone. I tried 22000, then 44100 and 48000 and at 64000 I got the best sound. My window size is 512 samples.
The microphone is a PDM microphone not PCM, but I'm told that the I2S driver interface takes care of that weird format and all I need to do is just read. I couldn't find any useable documentation on any of that - but I found a demo someone else had done also using 64000 and used the same settings. Whether or not PDM leads to different effective sample rates than PCM I have no idea. But the fundamental frequency emitted from ardiuinoFFT when I use a 1kHz tone is 1.000Khz. So - guess it must be correct.