A comparison of three noise reduction procedures applied to bird vocal signals

ABSTRACT.   Recordings of avian vocal signals in natural habitats include ambient noise. Often this background noise corrupts across all frequencies and is of substantial amplitude. Reducing this ambient noise to prepare vocal signals for playback stimuli or to remove habitat-specific noise signatures prior to analyzing a signal's acoustic characteristics can be useful. We conducted experimental evaluations of three noise reduction procedures to determine their effectiveness. We embedded two bird vocalizations ("clean" signals) in four kinds of natural noise, resulting in eight noise-signal combinations. We then applied three noise reduction procedures (Noise Profile, Band Pass, and Noise Estimate) to each of the embedded signals and compared the recovered signals to the original (clean) signals. Noise Profile filtering was effective in reducing noise and returning fairly high-quality signals from even severe levels of masking noise. The other two noise reduction procedures did not perform as well. For the two most corrupting maskers, however, Noise Profile filtering also altered the signal properties by reducing signal amplitude at those frequencies containing high levels of noise. Apart from this loss of amplitude, the quantitative features of the filtered signals were similar to those of the original model sounds. We conclude that Noise Profile filtering produces good results for cases where noise is approximately constant over the signal duration and the signal intensity exceeds noise intensity over the frequencies of interest.