Matched Filter Implementation

Matched Filter Matched filters do an excellent job of identifying sound samples, so we decided to apply the method here to identify birdcall audio files. A matched filter searches for a sample clip, the filter, within a longer audio recording. Convolution compares the filter to the longer signal at each possible offset. The greater the maximum amplitude of the convolution result, the stronger the match. By having a different filter for each birdcall, we can search an audio file to identify which birdcall it contains. The matched filter algorithm is as follows: Reverse the filters in the time-domain. Normalize the energy of each of the filters. Convolve each filter with the input signal and take the maximum amplitude of the resulting convolution signals. The filter that gives us the greatest maximum value indicates which birdcall the signal contains.

Filter Library Creation Our first step in implementing the matched filter algorithm was to create a library of birdcall filters. To do this, we looked at the spectrograms of a few sample audio files of the same birdcall and selected a portion that looked representative of the call.

For each birdcall, one of the representative audio segments was saved as a filter. Because the first two steps of the above matched filter algorithm affect only the filter library and are independent of the input signal, we reversed the filters and normalized their energy before saving them to wave files.

Matlab Implementation The following MATLAB script performed our matched filter algorithm. When given a wave file as input, it would tell us how well the audio sample matched against each of the 6 birdcall filters. We used circular convolution instead of linear convolution because it computed much faster. MATLAB's built-in cconv function zero-pads the two signals before multiplying their FFTs, generating the convolution result we are looking for.

function result = birdcheck(file)
    [sig, fs, nbits] = wavread(file);
    signal=sig(:,1);
    signal=signal/max(abs(signal));
    filters{1}=wavread('filters/bob.wav');
    filters{2}=wavread('filters/lt.wav');
    filters{3}=wavread('filters/lw.wav');
    filters{4}=wavread('filters/pygmy.wav');
    filters{5}=wavread('filters/red.wav');
    filters{6}=wavread('filters/redcry.wav');
    for i=1:6
      filter=filters{i};
      result(i) = max(abs(cconv(signal,filter(end:-1:1))));
    end 
  end

The script was able to correctly identify several birdcalls. It did fail to correctly identify four cases in two categories: Two of our loon tremolo files registered as pygmy owl common songs. Two of our red-tailed hawk shriek files registered as red-tailed hawk cries.