User's Manual - Cornell Lab of Ornithology - Cornell University

Chapter 7: Correlationsimilar. For example, in one set of 12 signals that were correlated in all 78unique pairs (including the 12 autocorrelations), the 22 correlations whosevalues exceeded 0.5 when using log spectrograms were ranked in the sameorder when using quadratic spectrograms from the same signals. Below thiscorrelation level, the exact rankings differed somewhat. Thus whether a givenpair of spectrograms has a higher correlation than another pair can depend inpart on whether log or quadratic spectrograms are used. You should thereforeconsider carefully whether logarithmic or quadratic spectrograms are moreappropriate for correlations given your application, and experiment with bothtypes if there is any question about which to use.Clipping levelSpectrogram correlations are calculated on the basis of all of the informationin a spectrogram, including noise. Sources of noise in a spectrogram includeacoustic noise in the original recording before it was digitized (e.g., fromwind or other environmental sources), and noise introduced as an artifact ofthe limited precision and dynamic range of the digitizing process, asdiscussed in Appendix A and Chapter 3.You can reduce the amount of noise in a logarithmic spectrogram by settingan appropriate clipping level when the spectrogram is made. To determine anappropriate clipping level, first make a spectrogram with a low clipping level,and then use the amplitude cursors in the spectrum pane to adjust the noisefloor of the spectrogram display upward, as explained in Chapters 1 and 3. 1Note the noise floor setting at which your signal shows up clearly withminimal noise visible on the spectrogram. Finally, recalculate the spectrogramusing this noise floor as the Clipping Level parameter.For normalized correlations, any change to the clipping level of the inputspectrograms (in either direction) can either raise or lower the peak correlationvalue, depending on the signals being correlated. For non-normalizedcorrelations, raising the clipping level always reduces the values of thecorrelation function.Grid resolutionThe size of the time increment between successive points in a spectrogramcorrelation function equals the time grid resolution of the inputspectrograms.The calculation time required for a spectrogram correlation is directlyproportional to the number of points in the spectrograms, and thus dependson the time and frequency grid resolution.1 If there is initially no noise visible as gray regions in the spectrogram that areobviously not part of the signal, try a lower clipping level.Canary 1.2 User’s Manual 119