Book excerpt: Abstract: "The problem of adapting the parameters of a speaker-dependent speech recognition system to a different speaker is examined with the objective of reducing or eliminating recognizer training necessary for user enrollment. A statistical approach to speech recognition based on vector quantization (VQ) and hidden Markov modeling (HMM) of speech is considered. The emphasis is on adaptation of vector quantizer prototypes as opposed to modification of hidden Markov model parameters. Two statistical techniques for VQ prototype adaptation, namely Bayesian learning and tied-mixture continuous-parameter HMM's, are presented and evaluated on the basis of experimental evidence. It is concluded that whereas Bayesian adaptation offers the best compromise between performance, amount of training data, and computational expense, tied-mixture continuous parameter HMM's constitute an even more reliable and effective technique for speaker adaptation."