This work examines the use of speaker classification as a method of improving speech recognition. Basic speech recognisers based upon hidden Markov models and neural networks, are modified by the use of selective training. Speakers are clustered into speaker types and separate recognisers are trained for each type. Performance is shown not to improve significantly. A more individualistic system is proposed. The speaker space is mapped by the use of non-linear interpolation between speaker dependent recognisers. Performance using an abstract 'perfect' speaker classifier is shown to be significantly better than speaker independent recognition. A multi-layer perceptron based speaker classifier is introduced, but is shown to be unable to learn the mapping from speakers to recognisers.
Bibliographic reference. Baldwin, David O. / Meyer, Georg F. (1995): "Improving speech recognition using speaker classification", In EUROSPEECH-1995, 1651-1654.