Accent adaptation in speech user interface

This paper examines the impact of accent present in speech on the performance of speaker independent automatic speech recognition (ASR) systems. In this paper, we show that, the presence of accent in the speech can increase the error rate. We validate a fundamental assumption that a speaker independent ASR engine, trained by a variety of accents, performs poorer than an engine that is trained for a particular accent, when tested by the same accent. Based on the results, we propose a method to lower the recognition error rate and measure the improvement by first determining the accent of the utterance and then applying the appropriate ASR engine from a bank of engines trained for different accents. We show that applying this method, will results to an average decrease of 24% overall error rate. The results are encouraging for a future complementary work. The research was carried out on an HMM based speech recognizer and TIMIT database was used to train and test the ASR engine.