Neural net approach for speaker sensitive measure analysis

Speech signals can be described by filters and excitation. This paper presents an analysis of the speaker information from this view point. We use the self organizing map of Kohonen (SOM) to explore the effectiveness of these two parts of a parameter in representing individual features of speakers. LSPs can be viewed as all-pole filters. Then, the excitation sequences of the filters are studied for discriminative characteristics of the speaker. In the experiment, three kinds of parameters-direct MFCC, LSP, and residual MFCC are used to build a feature map. Finally SOM shows that direct MFCC and LSP produce very similar feature maps for the same speaker in their general feature space. A correlation criterion gives further verification. While the SOMs of excitation signals are less discriminative between each other, they are less speaker sensitive.