Broad Phoneme Class Specific Deep Neural Network Based Speech Enhancement
Pavan Karjol, Prasanta Kumar Ghosh · 2018
In this work, we present a speech enhancement algorithm using broad phoneme class specific deep neural networks (DNNs). We build a classifier network to obtain the probabilities of each class in every test frame. The enhanced speech is obtained as the linear combination of the outputs from these class specific DNNs with combination weights as the probabilities predicted by the classifier network. We experiment with two (vowel and non-vowel) and four (vowel, stop, fricative, nasal) broad phoneme classes. Experiments are performed using speech data from TIMIT corpus, nine noise types (four seen and five unseen), and four SNR conditions. Experimental results show that the proposed class specific DNN approach outperforms the single DNN based speech enhancement scheme for both seen and unseen noise types in all SNR conditions considered.