A Low Complexity Algorithm for Isolated Sound Recognition using Neural Networks

This paper investigates a novel feature extraction algorithm suitable for low complexity implementations of sound recognition systems. The novelty consists in applying a simple method for generating feature vectors based on analysis of the lengths of blocks of identical consecutive significant bits of the signal sequence. Moreover, the above technique is applied to several consecutive signal windows from the speech signal, thus including the temporal features. Classification and recognition performances were evaluated on a database with 9 different users speaking 9 different sounds, each for 7 different instances. Results show similar performances to those obtained with more sophisticated methods such as the HMM method. For similar quality our method has a complexity reduced with 2 orders of magnitude, making it suitable for low-power, mobile applications.