The ability to accurately determine a sound source direction using standard signaling processing methods is difficult to accomplish, especially in reverberant environments. However mammals determine the origin of sounds without difficulty. The work presented here takes a "biomimetic" approach, to design an accurate sound source localization system that fulfills a need for efficient and compact sensor solution. Following previously developed biological models and signal processing methods of the mammalian hearing system, a real-time biomimetic digital system is developed that can be implemented on a single ASIC.

To achieve real-time results, the biomimetic processing model relies on large amounts of simultaneous computations: The acoustic data split into numerous parallel frequency channels, is processed through non-linear stages extracting relevant temporal cues that are encoded as "neuronal spikes". In past hardware implementations, such processing came at the detriment of area and power consumption. Therefore in the design presented, novel techniques at the architecture and arithmetic level allow minimized area and hardware overhead without forgoing system accuracy. Rather than having dedicated hardware on a per-frequency-channel basis, a specialized core channel, shared by all frequency bands is used. This time-shared architecture offers many other advantages, specifically in terms of system re-configuration (a feature not feasible until this point): processing coefficients, as well as the number and span of frequency channels, may be changed to fulfill various requirements and adapt to different applications without the need of system redesign. The nature of the proposed system is such that it may be easily adapted to fit a wide variety of applications ranging from hearing aids and machine monitoring to vehicle tracking and detection of small arms fire, mortars and heavy artillery.

The acoustic signal processing stages of a sound-source direction finding system were fabricated on a single ASIC using a 0.35um CMOS process. The overall area of the implementation occupies 12.44mm² and the processing of eight virtual frequency channels requires a clock speed of 30MHz while consuming 165mW. The system input sampling frequency being 100 kHz, assuming a microphone separation of 12 inches, the system achieves an azimuthal resolution of about 1 degree.