Speaker Identification of Whispering Sound using Selected Audio Descriptors

摘要

Whispering speech mode is adapted by speakers for any one of the reasons like secrecy of confidential data, avoiding being overheard in public places or hiding the identity intentionally. As acoustic properties of whispered speech are drastically changed compared to neutral speech; it makes difficult to identify the speaker from a whispered sound. This task requires the perceptual analysis of the whispering sound signal as do the humans. Many researchers presented various techniques for speaker identification of whispering sound but have some limitations. This paper describes the efficient method of identifying the speaker within the whispered speech using timbrel features which haven't been used so far in the whispered case. They are suitable here due to their multidimensional nature and perceptual ability. But all the timbrel audio descriptors are not well-performing for the whispered data. Hence, by using Hybrid Selection method, the most suitable timbrel audio features are selected and used. Timbrel features show an increase in the identification accuracy as 10.9% compared to traditional MFCC features. A database containing 650 utterances (whispered and neutral) of 35 speakers is created and used for the experiments. K-means classifier with a random choice of the centroid is used for classification.