ISCA Archive Interspeech 2016
ISCA Archive Interspeech 2016

Novel Subband Autoencoder Features for Non-Intrusive Quality Assessment of Noise Suppressed Speech

Meet H. Soni, Hemant A. Patil

In this paper, we propose a novel feature extraction architecture of Deep Neural Network (DNN), namely, subband autoencoder (SBAE). The proposed architecture is inspired by the Human Auditory System (HAS) and extracts features from speech spectrum in an unsupervised manner. We have used features extracted by this architecture for non-intrusive objective quality assessment of noise suppressed speech signal. The quality assessment problem is posed as a regression problem in which mapping between the acoustic features of speech signal and the corresponding subjective score is found using single layer Artificial Neural Network (ANN). We have shown experimentally that proposed features give more powerful mapping than Mel filterbank energies, which are state-of-the-art acoustic features for various speech technology applications. Moreover, proposed method gives more accurate and correlated objective scores than current standard objective quality assessment metric ITU-T P.563. Experiments performed on NOIZEUS database for different test conditions also suggest that objective scores predicted using proposed method are more robust to different amount and types of noise.