A Noise Reduction Method Incorporating Consonant and Vowel Characteristics for Dysarthric Speech Recognition

In this paper, a noise reduction method is proposed which incorporates consonant and vowel characteristics for dysarthric automatic speech recognition (ASR). Due to the noise-like acoustic characteristics of unvoiced consonants, Wiener filtering approaches may provide more distorted spectra of unvoiced consonants of dysarthric speech than those of voiced consonants. Thus, the proposed method selectively applies a Wiener filter or a Kalman filter depending on the voiced or unvoiced classification of consonants, respectively. In order to demonstrate the effectiveness of the proposed noise reduction method, ASR experiments are carried out on a database of mild and mild-tomoderate dysarthric speeches under different noise conditions. Consequently, it is shown that the proposed noise reduction method achieves relative average word error rate reductions of 20.45% and 8.10% for the mild and mild-tomoderate dysarthric groups, respectively, compared to that using a Wiener filter.