An Optimum Inexact Design for an Energy Efficient Hearing Aid

Abstract

Inexact design has proved to be an efficient way of obtaining power savings with a marginal penalty on performance in applications which do not need high degree of output accuracy. Such applications are often found in domains which deal with human sensorial systems. The 'not so perfect' state of human senses like sight, hearing which are important for video and audio related appliances can compensate for the error introduced due to inexact design. An efficient analysis and modelling of these compensation capabilities of human senses can help the designers to build optimum inexact architectures meeting the redefined requirements with low power. In this work, we choose hearing aid as our test application, which is based on human sense of hearing. We estimate a metric Intelligibility, for a set of audio samples, which is obtained from multiple surveys on human subjects to model the sensorial processing. Our methodology uses a novel way of introducing inexactness in an optimum manner. This includes a fine grained analysis of the error that is being introduced in the FIR filters of the DSP present in hearing aid. The resulting inexact FIR filter bank is 1.92× or 2.56× more efficient in terms of power consumed while producing 10% or 20% less intelligible speech respectively when compared with a hearing-aid using exact filters.

Evolutionary Algorithms in Approximate Computing: A Survey

“…And finally, in this category, we present two more sophisticated approaches. In [22], the parameter optimization is combined with constrained pruning. The objective is to approximate a digital signal processing block of a hearing aid consisting of m FIR filter banks.…”

Edge processing in IoT using approximate and in-memory computing

“…The energy requirement of the system can be brought down by approximate computing, where slight errors in calculation either do not change the outcome or can be ignored by human perception. Approximations in the computation are introduced from the basic building block of a circuit such as an adder [80], [116], or multiplier [3], [117] to the system level [118], [119]. Since the neural network is a good approximator, in this work, we have introduced approximation in the calculation at the system level using dynamic network expansion (DNS), where we have dynamically varied the number of neurons in the system based on anomaly detector decision.…”