Speaker age detection using eigen value

Abstract EN

In this research an algorithm was suggested for classifying a speaker age to two classes : (young and old classes) based on his speech signal.

The suggested algorithm depend on a speech signal feature extraction in order to get a compact representation for this signal and to adopte these features in the classification process.

In this algorithm, the eigen values of the covariance matrix was adopted as a principle parameter in the recognition between the two classes.

It was constructed from the data of the speech signal (usually one dimension) after rearranghng it into a number (2, 4, 8, 16, 32, 64) of a two dimension square matrix array.

The suggested algorithm include two main stages : 1st stage : includes data file preparation that contains the eigen values for a number of persons belong to both classes young and old (with different gender), in this stage the average of these values for each class to be calculated separately and the threshold curve(which represents the boundary seperating between two classes) were also computed.

Second stage : in this stage the classification process was done by comparing the curve that represents the eigen values of the speech signal, with the threshold curve, a different number of performance parameters are adopted in the evaluation the accuracy of the classification process.

The measured correlation value was in range of (0.9610, 0.9994) when m = 2 and m = 64, respectively (this means whenever the number of arrays that the speech signal constructed from it increases, the correlation coefficient also increases), while a clear difference can be seen with mmse parameter.

After applying the suggested algorithm on 50 persons from both genders, the algorithm passed in applying 80 % and failed in percentage 20 % of them.