Environmental Noise Robustness for Korean Fricatives using Speech Enhancement Generative Adversarial Networks

초록

Currently, speech recognition technology has a high recognition rate in a quiet environment condition. However, noise processing is needed in environmental noise conditions because they lower the recognition rate. In particular, in the case of Korean fricatives corresponding to [s], [c], [h], [c], [x], and [.], noise processing is relatively difficult because the acoustic characteristics are similar to environmental noise. This paper proposes speech enhancement of Korean fricatives using the Speech Enhancement Generative Adversarial Network (SEGAN) for environmental noise robustness. SEGAN is a Deep Neural Network (DNN)-based generation model that is used to improve the clarity and quality of environmental noisy speech. Enhanced Korean fricative speech data are generated using SEGANtrained Korean fricative speech data. The results showed that using DNN-Hidden Markov Model (HMM)-based acoustic model training by adding enhanced speech as training data resulted in a 0.26% Character Error Rate (CER) in environmental noise conditions compared to the case without enhanced speech addition.

키워드