Analog Lab: Computer Audition
We are designing custom integrated circuits for computer audition sensors. These are sensors that do not simply record sounds, but analyze and recognize specific acoustic signatures. They can be used for security, environmental monitoring and body-worn auscultation. A major inefficiency of current sensors is that they waste energy on acquiring, conditioning and digitizing data that is ultimately discarded by the DSP or some other downstream processor. We are developing intelligent, self-adaptive analog circuits that expend energy according to how computationally-relevant the data in question is.
A biologically-inspired nonlinear circuit determines the computational relevance of input sound. The noise floor, bandwidth and sampling rate of the primary signal processing chain (preamplifier (PA), antialiasing filter (AAF) and analog-to-digital-converter (ADC)) are adapted according to the information content of the input sound.
|Original, noisy input sound||Sound processed with our nonlinear algorithm, with the relevant regions preserved and the extraneous background noise discarded.|