Compact Isolated Speech Recognition on Raspberry-Pi based on Reaction Diffusion Transform

A low complexity solution for speech recognition is proposed and its implementation on a resources constrained platform, namely the Raspberry-Pi is evaluated. In order to achieve good performance with limited resources, both the feature extractor and the classifier are specially designed. A special form of feature extractor, called RDT (reaction-diffusion transform) was optimized and evaluated in conjunction with a specially designed ELM (extreme learning machine) classifier. Optimization of parameters led to very good recognition rates (up to 100%) for a small dictionary of isolated sounds representing vocal commands for automotive applications. The real time factor is sub-unitary, ensuring the realization of real time identification using the proposed method.