Low-complexity feature extraction unit for “Wake-on-Feature” speech processing

In the context of energy-constrained automatic speech recognition, a modeling and simulation tool is presented to evaluate the hardware complexity of a feature extraction unit. The objective is to evaluate the minimal amount of features necessary for voice-activity detection, considering limited hardware resources. The obtained features are fed into a classification engine to evaluate the ability to differentiate human voice from background noise. While keeping a detection accuracy of passwords from a standard data-set at 90%, the study shows that the parameters of the feature extraction components, such as ADC resolution or energy quantization resolution, can be reduced to 8 bits and 6 bits, respectively, considering 8 frequency bands, in order to allow hardware-efficient implementations.