This paper presents a new prediction error normalisation method for speaker verification using predictive neural networks. A prediction error obtained by a predictive neural network strongly depends on a particular input and the goodness of the fit is difficult to determine by comparing the value with a fixed threshold. We propose a prediction error normalisation algorithm which uses the prediction error obtained by a network trained for multiple categories as a measurement of predictability of an input.
The algorithm was evaluated in text-independent speaker verification. Without normalisation, an equal error rate of 41.2% was achieved for 12 male speaker verification, using the normalisation, the equal error rate was improved drastically to 1.5%. This result proved the effectiveness of the proposed algorithm.
The proposed algorithm is also applicable to other speech processing areas which involve comparison with a threshold such as word spotting and rejection of unknown words.