A comprehensive computational model of animal biosonar signal processing
Fig 3
The input is an FM pulse followed by an echo at delay (t) with two glint reflections separared by (Δt). These signals are segmented into parallel frequency channels by 161 bandpass filters (4th order gammatone filters, center frequencies from 20 to 100 kHz, equivalent rectangular bandwidth, ERB, 4 kHz), then half-wave rectified and lowpass filtered (10 kHz cutoff). Outputs are then normalized to fit into 10 equally spacedthreshold levels (3% to 98% full-scale). Processing started at the lowest frequencies in the broadcast in bats 23–25 kHz); if these particular frequencies are absent, then echoes are rejected and processing is not started [64]. To remove the slope of the FM sweeps, echo frequencies are dechirped on the time-frequency plane by setting zero origin of time for individual bandpass frequency channels to threshold times of the pulse, resulting in corresponding thresholds of echoes arrayed at longer delay corresponding to target range. These threshold times are then analyzed by two parallel pathways—range delay from total echo delay (t) determined by time between pulse and echo thresholds combined across all broadcast frequencies, and glint delay (Δt) from spectral notch detection followed by extraction of time spacing for glint reflections before merging of range delay and glint spacing onto a common perceived delay scale (blue, at output).