Robust Phoneme Recognizer at Noise Corrupted Acoustic Environment

This paper proposes a robust automatic phoneme recognizer for Japanese language in noise corrupted acoustic environments. From the previous studies it is found that some hidden factors such as speaking style, gender effects, and noisy acoustic environments degrade the performance of automatic phoneme recognizers (APRs). In this study, an APR is designed in noise corrupted acoustic environments resolving the noise effect. The proposed system comprises three stages. At first stage, a multilayer neural network (MLN) that outputs Distinctive Phonetic Features (DPFs) from the input acoustic features is incorporated, and then the Karhunen-Loeve Transformation (KLT) and the Gram-Schmidt (GS) algorithms are used at second stage to extract reduced feature vector. Finally, the output phoneme strings are generated by inserting the reduced features into a hidden Markov model (HMM) based classifier. It is observed from the experiments in clean and noisy acoustic environments that the proposed method provides higher recognition accuracy at lower Signal-to-Noise Ratios (SNRs).