Andrew T. Sapeluk

Impostor cohort selection for score normalisation in speaker verification

Abstract Impostor cohort normalisation (ICN) may be used to improve speaker verification performance. The choice of impostors for the cohort is critical to the success of ICN. To select suitable impostors, they must be ranked according to their ability to impersonate the genuine speaker. In the technique described here, a genuine speakers impostors are ranked according to how well their models score against the genuine-speaker model. This is faster than the traditional method of testing the model against the impostor training utterances, and is in good agreement with it. Having obtained rankings, impostors were selected in three different ways: of these, cohorts based on the closest impostors gave the best results, reducing the unnormalised equal error rate by more than half.

Finite element aided design optimisation of a shaded-pole induction motor for maximum starting torque

A method of optimizing the design of a shaded-pole induction motor for maximum starting torque with the aid of finite element modeling and a modified hybrid global-local search method combining the niching genetic algorithm with a direct search method is presented. By invoking the genetic algorithm and the deterministic method in turn, the solution with the global minimum is secured while simultaneously improving the convergence speed. The performance of the hybrid search method is demonstrated with an ideal mathematical problem first, before applying it to the shaded-pole motor design.

Improved Data Modeling for Text-Dependent Speaker Recognition Using Sub-Band Processing

A growing body of recent work documents the potential benefits of sub-band processing over wideband processing in automatic speech recognition and, less usually, speaker recognition. It is often found that the sub-band approach delivers performance improvements (especially in the presence of noise), but not always so. This raises the question of precisely when and how sub-band processing might be advantageous, which is difficult to answer because there is as yet only a rudimentary theoretical framework guiding this work. We describe a simple sub-band speaker recognition system designed to facilitate experimentation aimed at increasing understanding of the approach. This splits the time-domain speech signal into 16 sub-bands using a bank of second-order filters spaced on the psychophysical mel scale. Each sub-band has its own separate cepstral-based recognition system, the outputs of which are combined using the sum rule to produce a final decision. We find that sub-band processing leads to worthwhile reductions in both the verification and identification error rates relative to the wideband system, decreasing the identification error rate from 3.33% to 0.56% and equal error rate for verification by approximately 50% for clean speech. The hypothesis is advanced that, unlike the wideband system, sub-band processing effectively constrains the free parameters of the speaker models to be more uniformly deployed across frequency: as such, it offers a practical solution to the bias/variance dilemma of data modeling. Much remains to be done to explore fully the new paradigm of sub-band processing. Accordingly, several avenues for future work are identified. In particular, we aim to explore the hypothesis of a practical solution to the bias/variance dilemma in more depth.

VQ Score Normalisation for Text-dependent and Text-independent Speaker Recognition

Individual weighting of speaker models in VQ-based recognition has some advantages but means that scores from different models may not be directly comparable, so making identification difficult. It is also problematic for verification as decision thresholds cannot easily be set without first testing models with genuine and imposter utterances. We present a novel normalisation method for VQ speaker recognition which applies an offset to each model, based on the average score between it and the imposter models, to bring particularly high- or low-scoring models into line with the general score range. It may be calculated a priori, before running any actual tests. The method works for both text-dependent and text-independent tasks and improves both the identification and verification error rates.

A model to support the authentication of mobile business

Identity authentication and access control are the most important aspects of internet security. The traditional method to implement them is achieved through an API between the authentication and access control service provider with an application strengthened with one-time password. These methods do not meet the security demands and a new identification mechanism is required. We present a multilevel smart authentication mechanism which defines four different alert levels in the system according to users requirements, using three different authentication techniques. It will not only introduce systematic safety, but also more convenience to the user.

An investigation into using neuro-evolution of Augmenting Topologies (NEAT) for short term load forecasting (STFL)

Different ANN architectures using back propagation can forecast the electricity demand at half-hourly intervals for up to 24 hours ahead with various degrees of success that is highly dependent on mainly trial and error heuristic tailoring of the architecture and the various learning parameters to cover the solution space. This paper presents the results of an investigation of an approach in the neuro-evolution technique to the short term electricity forecasting (STFL) problem. This algorithm is called Neuro-Evolution through Augmenting Topologies (NEAT). We have chosen the methodology in the paper to be a simple, generic, adaptive, robust, and easy to implement approach, requiring modest computing resources, for the prediction of the electricity demand.