Development and Testing of a Compact Voice Command Recognition Algorithm for Limited Complexity Microcontroller Devices

Abstract

We describe and test an effective language-independent voice command recognition algorithm taking into account the possible home and industrial automation tasks and cost effective microcontroller and/or single board computer realizations. The algorithm is based on flexible time warping of spectrogram tables and statistical fuzzy logic processing of in-group and out-group discrepancy parameter typical values. The new key parameters introduced are the discrepancy [dB/square] between every two voice commands and rejection-recognition match parameter between −1 and +1 to characterize probability that the command-under-test belongs to a certain group. The developed algorithm complemented by several bonus and penalty mechanisms made possible to reach 97%-99 % recognition rate with a steep learning curve demanding only 10 example commands per group. Further testing also showed that the algorithm was capable to demonstrate a reasonable 80%-85 % accuracy with only 1–3 example commands.