Combining Genetic Algorithms and FLDR for Real-Time Voice Command Recognition

In this article we propose an effective method for a user-dependant voice command (small vocabulary) recognition system based on the combination of genetic algorithms and the Fisher’s Linear Discriminant Ratio (FLDR). A genetic algorithm here is used to search in the frequency domain of the voice for those sub-bands whose energy is discriminant enough so as to distinguish between at least two different classes, and at the same time, being able to appropriately agglomerate the different utterances of one word of the vocabulary in a compact class. Once the sub-bands have been evolutionary selected, its energy  is represented in feature vectors; in this way, very few samples of each voice command are required to build each word's model; moreover, this is a convenient method for feature selection. Real time implementation was done in a DSP TMS320LF2407 using elliptic bandpass filters -– one per sub-band -- with floating point representation. Encouraging results were obtained.