A text independent speaker identification system using signal processing techniques and artificial neural networks

Abstract

Identifying a speaker by the speech signal uttered by the speaker is very essential in many applications, such as voice controlled applications, fraud call detection, information retrieval systems, etc. The speakers have to be recognized irrespective of the content they speak. This is a challenging area which has good scope for research. Speaker recognition is very popular in various present day fields which include authentication of a person such as newlinesecurity control, tele-banking, etc. This is highly essential when authentication have to be done remotely, because the other forms of identification requires physical contact of the person. In this thesis work, a text independent speaker recognition system is developed, which extracts robust features from the speech signals and classifies them with good accuracy. Initially, the speech signals are preprocessed which removes the silence portion in the signal, emphasizes the high frequency components and reduces the noise. In the first method, temporal features such as Energy, Mean, Variance, Skewness and Kurtosis are extracted from the speech signals. These features are selected as they analyze the amplitude content in the input signals. Five different databases are considered, namely Texas Instruments and Massachusetts Institute of Technology (TIMIT), Free Spoken Digit Database (FSDD), English Language Speech Database (ELSDSR), Student voice Database (STVOICE) and Telephone Voice Database. Out of these, the first three are standard databases and the other two are realtime databases. newline newline