Analytical review and classification of methods for features extraction of acoustic signals in speech systems

This paper presents an overview of methods and algorithms for feature extraction to transform an acoustic signal into a sequence of vectors for solving problems of segmentation, classification, identification, or speech recognition. A classification of feature extraction methods according to mathematical approaches is proposed. The algorithms and techniques of spectral analysis, which are most used in the design of speech recognition systems, are discussed. This review clearly demonstrates the complexity of the problem of acoustic processing - searching a representation that decreases the dimension of the model and maintain the completeness of linguistic information and, importantly, is stable to variability with respect to the speaker, transmission channels and the environment. The analysis of the existing feature extraction methods is useful for selection of a technology when designing a key element of a speech system.