Figure 3
Block diagram for one chamber showing the digital signal processing blocks on the FPGA (blue) and analog devices (black). To compose the Speaker signal (green, left) destined to the loudspeaker, we selectively added signals from other chambers with playback and noise signals that we then band-pass (BP) filtered. Because the microphone picks up the signals from both the bird (red) and the loudspeaker (green), we subtracted from the Mic signal (purple) the speaker component (SpeakerEcho) that we estimated with a least mean square (LMS) adaptive filter, resulting in the separated bird signal in the chamber (MicSep). A subsequent squelch passed the MicSep signal only when its intensity was above a threshold. The threshold was the sum of a constant value and a dynamic part that was proportional to the intensity of the SpeakerEcho signal. The dynamic threshold provided robustness to unwanted sound leakage from an imperfect echo attenuation with the LMS filter. To avoid cutting vocalization onsets, we introduced a delay of the MicSep signal into the squelch. The resulting chamber output signal was the separated and squelched microphone signal (MicSepSq). Indicated by curved arrows are the gain \({\mathrm{g}}_{\mathrm{o}}\) of the output chain, the gain \({\mathrm{g}}_{\mathrm{i}}\) of the input chain, and the chamber gain \({\mathrm{g}}_{\mathrm{c}}\).
