Page 21 - Acoustics Today

Fig. 2. Theoretical on-axis gain of conical and exponential horns relative to a free- field microphone. Two lengths are represented: 88 mm for fc = 2777 Hz, and 44 mm for fc = 5554 Hz.
an effective plan for processing the data. In the authors’ expe- rience, many projects have encountered significant bottle- necks in this phase. There are many potential software tools to automate processing, but the difficulty seems to lie in selecting the appropriate tool, and gaining sufficient experi-
ence with its use to maximize its performance.
Automated bioacoustical monitoring can be envisioned
as a three step process: detecting and delimiting events of interest, characterizing the structure of the event, and classi- fying the event to a species or other class of signals. In some processing schemes, a highly selective detector serves all three processes. Examples are matched filtering17 and spec-
18
Matched filtering is an optimal detec- tor when the signal is known exactly, a situation that will rarely obtain for biological sources. In fact, it is not an exag- geration to suggest that no animal sound is ever replicated exactly. The limited time-frequency resolution of spectro- grams helps diminish small differences among acoustical sig- nals, so spectrogram correlation is not as narrowly selective as waveform matched filtering. Nonetheless, the authors’ experience has shown that successful use of spectrogram cor- relation for some classes of sounds may require tens of tem- plates to achieve adequate coverage. Less selective detectors19 defer the decisions about classification, and offer researchers opportunities to study a wide range of signals that share some structural similarity with the signals of interest. This flexibil- ity comes at a cost—the detection software may need to man- age millions of detections, and subsequent analysis will chal-
trogram correlation.
Table 3. Web links to software for bioacoustical data processing
20 Acoustics Today, July 2012