FPGA Implementation of a Configurable Vocal Feature Extraction Embedded System for Dysarthric Speech Recognition

In the last few years, users have been increasingly demanding for a hands-free interaction with their digital devices. This kind of technology is even more useful if used by people with disabilities, improving their quality of life. In particular, speech-impaired users (e.g. dysarthric speakers) represent a big challenge for an Automatic Speech Recognition (ASR) system because standard approaches are ineffective with them. Therefore, new speech analysis algorithms are implemented and generally tested on off-line datasets, but their performance can differ from a real case. Hence comes the need to easily validate their performance in a real scenario. The work presented in this paper shows an implementation of a highly configurable off-line embedded system for both MFCC and Mel Filterbanks extraction equipped with an on-board microphone. The results show that our system performs well in a real scenario case in terms of both power consumption and word error rate.