Neural approach to voiced-unvoiced-silence analysis for quality measurements in telecommunication systems

摘要

A new approach for classifying the speech signal into voiced and unvoiced speech is proposed. It is based on the segmentation of the speech signal in successive and separated frames. All frames are pre-processed and successively classified by means of the Learning Vector Quantisation neural network. The proposed approach is a key step in the design and implementation of an In-service Non-intrusive Measurement Device. Experimental tests showing the efficacy of the proposed approach are performed by using stationary noise, impulsive noise, and Modulated Noise Reference Unit. Test results are discussed and compared with the results obtained by different techniques presented in literature.