Affine transformations in speaker adaptation – why simpler is better

Speaker adaptation is an important technique that can compensate for the mismatch between training data and the vocal characteristics of an individual user in a speech recognition system, however this can come at the cost of increased computational complexity. This paper reports a detailed comparison of four different affine transformation configurations for speaker adaptation, and the evaluation of their recognition accuracy, complexity and memory requirements. Results of this comparison show that for optimum parameter choices, simpler transformation configurations are capable of producing accuracies close to those of the conventional full transformation, allowing the computational complexity to be reduced by one to two orders of magnitude.