Rapid Speaker Adaptation Research Articles

Combined Bayesian and predictive techniques for rapid speaker adaptation of continuous density hidden Markov models

One problem faced by some model adaptation techniques is that only the parameters of those models which are observed in the adaptation data are updated. Hence, with small amounts of adaptation data most of the system parameters remain unchanged. In this paper, a technique called regression-based model prediction (RMP), which tries to overcome this problem, is presented. This technique tries to adapt the model parameters of a continuous density hidden Markov model set which has insufficient adaptation data when used with maximuma posteriori(MAP) estimation. The technique uses the parameters of better estimated models and a set of parameter relationships between the model parameters to update the parameters of models with insufficient adaptation data. The parameter relationships are found by applying linear regression to a number of speaker-specific model sets. Experiments using both MAP estimation and RMP are presented using the ARPA RM1 continuous speech database and RMP has been found to be useful in improving the system performance with as little as 3 s of adaptation speech. RMP has been shown to consistently improve the results obtained by MAP. When a very large number of adaptation sentences are used the error rates converge towards those of MAP. It is shown that RMP gives an improvement of 8·8% over the baseline error rate with a single adaptation sentence, and 27% with 40 adaptation sentences.

Speaker normalization and adaptation using second-order connectionist networks

A method for speaker normalization and adaption using connectionist networks is developed. A speaker-specific linear transformation of observations of the speech signal is computed using second-order network units. Classification is accomplished by a multilayer feedforward network that operates on the normalized speech data. The network is adapted for a new talker by modifying the transformation parameters while leaving the classifier fixed. This is accomplished by backpropagating classification error through the classifier to the second-order transformation units. This method was evaluated for the classification of ten vowels for 76 speakers using the first two formant values of the Peterson-Barney data. The results suggest that rapid speaker adaptation resulting in high classification accuracy can be accomplished by this method.