Android: Finding fundamental frequency of audio input

Question

I've been trying for some time now to find the best solution to calculate the fundamental frequency of a sample captured using AudioRecord in real-time.

I have looked around some examples around here on SO: This one, and this one are the questions that helped me the most, but I still did not understand fully how they would work for finding the fundamental frequency. So what I am looking for is a more detailed explanation of what do I need to do to find the fundamental frequency having a sample.

So, I create an AudioRecord:

micData = new AudioRecord(audioSource, sampleRate, channel, encoding, bufferSize);
data = new short[bufferSize];

And start listening:

micData.startRecording();    
sample = micData.read(data,0,bufferSize);

And I understand how to create a Complex array, but I don't know exactly which methods out of FFT.java I can use the values of to create these complex numbers and the which one would be the method that returns the peak frequency.

Answer 1

Reading your question I see you are not sure yet that you want to use FFT. That's good because I don't recommend using just FFT. Stay in time domain, use Autocorrelation or AMDF and if you want more accurate results, than use FFT as a additional component.

Here is my Java code for calculating fundamental frequency. I wrote comments because you say you still don't understand the process.

public double getPitchInSampleRange(AudioSamples as, int start, int end) throws Exception {
    //If your sound is musical note/voice you need to limit the results because it wouldn't be above 4500Hz or bellow 20Hz
    int nLowPeriodInSamples = (int) as.getSamplingRate() / 4500;
    int nHiPeriodInSamples = (int) as.getSamplingRate() / 20;

    //I get my sample values from my AudioSamples class. You can get them from wherever you want
    double[] samples = Arrays.copyOfRange((as.getSamplesChannelSegregated()[0]), start, end);
    if(samples.length < nHiPeriodInSamples) throw new Exception("Not enough samples");

    //Since we're looking the periodicity in samples, in our case it won't be more than the difference in sample numbers
    double[] results = new double[nHiPeriodInSamples - nLowPeriodInSamples];

    //Now you iterate the time lag
    for(int period = nLowPeriodInSamples; period < nHiPeriodInSamples; period++) {
        double sum = 0;
        //Autocorrelation is multiplication of the original and time lagged signal values
        for(int i = 0; i < samples.length - period; i++) {
            sum += samples[i]*samples[i + period];
        }
        //find the average value of the sum
        double mean = sum / (double)samples.length;
        //and put it into results as a value for some time lag. 
        //You subtract the nLowPeriodInSamples for the index to start from 0.
        results[period - nLowPeriodInSamples] = mean;
    }
    //Now, it is obvious that the mean will be highest for time lag equal to the periodicity of the signal because in that case
    //most of the positive values will be multiplied with other positive and most of the negative values will be multiplied with other
    //negative resulting again as positive numbers and the sum will be high positive number. For example, in the other case, for let's say half period
    //autocorrelation will multiply negative with positive values resulting as negatives and you will get low value for the sum.        
    double fBestValue = Double.MIN_VALUE;
    int nBestIndex = -1; //the index is the time lag
    //So
    //The autocorrelation is highest at the periodicity of the signal
    //The periodicity of the signal can be transformed to frequency
    for(int i = 0; i < results.length; i++) {
        if(results[i] > fBestValue) {
            nBestIndex = i; 
            fBestValue = results[i]; 
        }
    }
    //Convert the period in samples to frequency and you got yourself a fundamental frequency of a sound
    double res = as.getSamplingRate() / (nBestIndex + nLowPeriodInSamples)

    return res;
}

What else you need to know is that there are common octave mistakes in the autocorrelation method especially if you have noise in the signal. From my experience, piano sound or guitar isn't problem. The mistakes are rare. But human voice could be...