Implementation of Automatic Speech Recognition for Low-Power Miniaturized Devices

This paper describes the development and implementation of a discrete speech recognizer for application in miniaturized, portable low-power devices. The recognizer is based on Hidden Markov Models (HMMs) and uses Linear Predictive Coefficients (LPC) converted to cepstrum to parameterize the input speech. The implemented recognizer has a complexity of 0.67 MIPS and a recognition performance of 97.69 % when recognizing isolated digits in French. The physical implementation of the recognizer was done on a proprietary ASIC DSP and verified using fast prototyping on FPGA.