Our goal has been to implement acoustic sensors based on biomimetic signal processing methods for sound source localization. Mimicking the mammalian auditory system, the sound localization is based on interaural time delay (ITD) estimates, which are converted to bearing.
Two generations of hardware using discrete off the shelf analog and digital components have been developed in our laboratory in the past. The first supported two microphones, each feeding 16 frequency bands. Each band fed spiking neurons having three different thresholds, analogous to those found in the mammalian auditory system. The overall size of the assembly was on the order of 300 cubic inches. The bandwidth was about 1 kHz and the system was used mostly for vehicle tracking. Next generation hardware supported four microphones, with again 16 frequency bands derived from each microphone. The volume of this device was approximately 150 cubic inches and achieved a bandwidth of approximately 10 kHz. This system was used chiefly as a sniper detector. Size reduction was achieved by using improved printed circuit board (PCB) design as well as miniaturized discrete elements.
The first attempt toward this goal focused on a fully digital implementation of the model that would allow further size reduction and hence result in a small size portable system. Our first step towards that goal was using digital signal processing (DSP) chips, in this case a TMS320C6713. We have implemented parts of our system and made estimates of the number of TMS320C6713 chips that should be required to implement a system comparable to those described above: Four chips are required to service four microphones of 16 channels each with a sampling frequency of 10 KHz. Our estimate of the volume containing those chips, each with related sub-circuitry, is on the order of 20 cubic inches.
The most recent stage of DSP miniaturization was carried out using Analog Devices, Inc. (ADI) TigerSHARC TS-201 chips. Per chip, we achieved a real time implementation of 14 frequency channels ranging from 300Hz to 10 KHz, each servicing three populations of spiking neurons at a sampling frequency of 96 KHz. This implies that the highest temporal resolution of the current implementation is approximately 10us time intervals.
We have focused our efforts on optimizing the prior software algorithm. Our optimization, to date, was successful in increasing the processing speed by six times compared to the initial software and ten times compared to the prior Texas Instruments TMS320C6713 based system. To service four microphones with capabilities noted above, an estimated four TS-201 chips are required. The approximate volume needed to implement a portable system including all required sub-circuitry, is on the order of 30 cubic inches; compared to the 150+ cubic inches of the developed analog boards.