We investigate speaker identification in narrowband noise using subband processing. The output of each subband is used to train and test individual hidden Markov models (HMMs), each making a preliminary decision on speaker identity. Subsequently, these are combined to produce a final decision. For sufficient numbers of filters, subband processing outperforms traditional wideband techniques by an enormous margin.

Previous work has demonstrated the performance gains that can be obtained in speaker recognition by applying subband processing, together with hidden Markov modelling and multiple classifier recombination. Two recombination rules have been investigated: the sum of log likelihoods, which corresponds to the optimal Bayes' rule under certain constraints, and multilayer perceptrons (MLP), which are not subject to these constraints. It was found that for two spoken digits in the presence of a single case of narrowband noise the sum of log likelihoods and MLP achieved comparable performance. In this paper, the previous work is extended in the direction of investigating the robustness of the recognition system to different narrowband noise. Two approaches are taken towards this aim. Firstly, narrowband noise is added at different centre frequencies. Secondly, a Bayesian MLP approach is investigated using automatic relevance determination (ARD) on the subband inputs to the MLP. From this it is possible to assess the relative importance of the subbands to recognition performance. Results for the new noise conditions show that the sum of log likelihoods generally does better than the (average) MLP fusion.

No abstract

No abstract