A novel feature extraction method for isolated word recognition based on Nested Temporal Averaging

Abstract

A novel preprocessing method is proposed. It has a reduced complexity and therefore is aimed to be used in low power, VLSI implemented, speech recognizers. Our algorithm extracts a feature vector made from up to 3 feature vectors, each coming from a particular variable length speech sequence. The sequences are nested one into each other while their length is divided by 2 for each nesting operation. Each feature vector is computed as an average, min and max of all 13-dimensional Mel-cepstral coefficients obtained within a sound sequence. On a sound database with 10 speakers speaking 7 different words the classification performance was found to be close and even better than the one obtained using traditional methods (HMMs).