The use of hybrid lvq3 neural networks for speaker recognition problems

Abstract EN

The main objective of this work is to develop the learning vector quantization (LVQ) neural network for the speaker recognition problems.

The proposed methods consist of two main stages: the analysis stage and the recognition stage.

The analysis stage was achieved by using wavelet transform coefficients, The neural networks was used in the recognition stage.

The Four speaker recognition systems were differently constructed to appropriate the coefficients results in the stage of analysis which it was achieved by discrete wavelet packet transform (DWPT) coefficients.

The first system was implemented by using the third development of LVQ (LVQ3).

In other words it is pure neural network.

The rest of the proposed systems were implemented by using different proposed activation functions of this network.

The second system was implemented by combining the three sigmoidal functions in one function .The third system was implemented by a tansh function, and the last system was implemented using a wavenet network under which the second derivative of Haar wavelet function will be used.

The work was applied on 10 persons (5 males and 5 females) which have converge ages.

The results of the test showed that recognition rate averages for the five-recorded speaker statement which not previously trained.