Improving automatic speech recognition robustness for the Romanian language

In this paper we propose an alternative way of improving speech recognition accuracy by analyzing the relevance of voice feature dimensions. A new measure is defined in order to quantify the feature distribution overlapping. Based on this measure, weights for voice feature dimensions are calculated and then applied to the hypotheses resulted from an N-best recognition process. Experiments are made with an Automatic Speech Recognition (ASR) system for the Romanian language. A relative improvement of 22% is obtained in terms of Word Error Rate (WER).