Speaker verification using heterogeneous neural network architecture with linear correlation speech activity detection

Abstract

This paper presents a multi-level speaker verification system that uses 64 discrete Fourier transform spectrum components as input feature vectors. A speech activity detection technique is used as a pre-processing stage to identify vowel phoneme boundaries within a speech sample. A modified self-organising map (SOM) is then used to filter the speech data by using cluster information extracted from three vowels for a claimed speaker. This SOM filtering stage also provides coarse speaker verification. Finally, a second speaker verification level of three multi-layer perceptron networks classifies the filtered frames provided by the SOMs. These multi-layer perceptrons work as fine-grained vowel-based speaker verifiers. The proposed verification algorithm shows a performance of 94.54% when evaluated using 50 speakers from the Centre for Spoken Language Understanding speaker verification database. In addition, it is shown that the novel discrete Fourier transform spectrum-based linear correlation pre-processing technique, presented here, provides the system with greater robustness against changes in speech volume levels when compared with an equivalent energy frame analysis.

Speaker verification inspired by the physiology of hearing using spiking self‐organising map

“…The three vowels are contained in the words (two, five and eight) of the CSLU2002 database. The position of the target vector for each neuron is chosen after locating the vowel region within each word in the enrolment speech sample using the pre-processing linear correlation technique presented in the work of Tashan et al (2012). When the spike timings of an input vector are fully synchronised with the spike timing of the target vector, each input synapse connected to the spiking neuron will respond with a maximum value of 1, resulting in an output of 1 as shown in Figure 8b.…”