Hardware Accelerator for HMM Based Speech Recognition using Approximate Computing Techniques

Abstract

This thesis presents a hardware design for recognizing speech using phoneme-level Hidden Markov Models (HMMs) and proposes two alternative designs using approximate computing techniques for area and energy optimizations.

An initial hardware design is proposed to model a speech recognition system using the log-Viterbi algorithm approach. Two more hardware designs using various approximate computing techniques and modifications to the log-Viterbi algorithm are also proposed, that are shown to consume lesser area and power.

The work also presents the performance analysis in terms of recognition accuracy and hardware evaluations in terms of area, switching and leakage power and energy dissipation of all three designs. The results prove that the usage of approximate computing helps reduce area and power, with a minor compromise on accuracy. The design using approximate computing is also capable of running at a higher frequency with quicker execution time and lesser energy consumption.

For applications where accuracy is vital, the thesis also proposes an adaptive system which can operate in two modes – one at a higher frequency, with slightly lesser accuracy and another at a lower frequency, with better accuracy and capable of dynamically switching from one mode to another.