Two-Stage Unsupervised Model Adaptation to Improve the Performance of Connected-Word Recognition

If HMH word models for automatic speech recognition have not been trained by the current speaker, the recognition performance can be improved by adapting such models during use to make them better represent the observed properties of the input speech.There are two quite distinct types of inadequacy in word models that may cause recognition errors.The first type is a result of a general difference in spectral properties of the speech of the current user from those of the speech used as training data.Such differences can be causedl for example, by different microphones and differences in typical glottal source spectrum.The second class of difference is specific to particular phonetic events, depending on articulatory detail of individual words.Correction of general spectral trends is much more robust than adaptation of individual p.d.f.s, so it is advantageous to spend the first minute or so of the adaptation correcting the spectral trend effects only.This process by itself should normally improve the recognition accuracy, thus increasing the effectiveness of subsequent unsupervised adaptation for modifying the individual p.d.f.s.Results are presented for several speakers using a simple connected word stochastic recognizer, comparing this two stage adaptation with simple adaptation of p.d.f.s only.