Improved Example-Based Speech Enhancement by Using Deep Neural Network Acoustic Model for Noise Robust Example Search

Example-based speech enhancement is a promising singlechannel approach for coping with highly nonstationary noise. Given a noisy speech input, it first searches in a noisy speech corpus for the noisy speech examples that best match the input. Then, it concatenates the clean speech examples that are paired with the matched noisy examples to obtain an estimate of the underlying clean speech component in the input. The quality of the enhanced speech depends on how accurate an example search can be performed given a noisy speech input. The example search is conventionally performed using a Gaussian mixture model (GMM) with mel-frequency cepstral coefficient features (MFCCs). To improve the noise robustness of the GMM-based example search, instead of using noise sensitive MFCCs, we have proposed using bottleneck features (BNFs), which are extracted from a deep neural network-based acoustic model (DNN-AM) built for automatic speech recognition. In this paper, instead of using a GMM with noise robust BNFs, we propose the direct use of a DNN-AM in the example search to further improve its noise robustness. Experimental results on the Aurora4 corpus show that the DNN-AM-based example search steadily improves the enhanced speech quality compared with the GMM-based example search using BNFs.