Acoustic factorisation

This paper describes a new technique for training a speech recognition system on inhomogeneous training data. The proposed technique, acoustic factorisation, attempts to model explicitly all the factors that affect the acoustic signal. By explicitly modelling all the factors, the trained model set may be used in a more flexible fashion than in standard adaptive training schemes. Since an individual model is trained for each factor, it is possible to factor-in only those factors that are appropriate to a particular target domain, for example the distribution over all training speakers. The target domain specific factors are simply estimated from limited target specific data, for example the target acoustic environment. The paper describes the theory of this new approach for the transforms for a particular speaker and environment. Initial experiments on a large vocabulary speech recognition task are presented.