A Clustering Algorithm for the Fast Match of Acoustic Conditions in Continuous Speech Recognition

Abstract

In practical speech recognition applications, channel/environment conditions may not match those of the corpus used to estimate the acoustic models. A straightforward methodology is proposed in this paper by which the speech recognizer can match the acoustic conditions of input utterances, thus allowing instantaneous adaptation schemes. First a number of clusters is determined in the training material in a fully unsupervised way, using a dissimilarity measure based on shallow acoustic models. Then accurate acoustic models are estimated for each cluster, and finally a fast match strategy, based on the shallow models, is used to choose the most likely acoustic condition for each input utterance. The performance of the clustering algorithm was tested on two speech databases in Spanish: SENGLAR (read speech) and CORLEC-EHU-1 (spontaneous human-human dialogues). In both cases, speech utterances were consistently grouped by gender, by recording conditions or by background/channel noise. Furthermore, the fast match methodology led to noticeable improvements in preliminary phonetic recognition experiments, at 20-50% of the computational cost of the ML match.