// For audio recognition, we need a double-precision (16 bit)
// integral image. We simply turn the sample data into pixels and
// have ourselves a verybig*1 image.
srecord noeq AudioRecognizer {
  interface IAudioSample {
    double length();
    double sampleSum(int channel, double start, double end);
    
    public default double getPixel(int channel, double start, double end) {
      ret doubleRatio(sampleSum, end-start);
    }
  }
  
  // the 16 bit per channel integral image
  class AudioSample implements IAudioSample {
    int channels;
    long[] data;

    public double length() { ret data.length; }
    
    public double sampleSum(int channel, double start, double end) {
      int a = ifloor(start), b = ifloor(end);
      ret getEntry(getEntry(b)-getEntry(a-1);
    }
    
    // get an entry of the sum table
    long getEntry(int i) {
      if (i < 0) ret 0;
      ret data[min(i, data.length-1)];
    }
    
    *(short[] samples) {
      int n = l(samples);
      data = new long[n];
      long sum = 0;
      for i to n:
        data[i] = (sum += samples[i]);
    }
  }
}