Robust Example Search Using Bottleneck Features for Example-Based Speech Enhancement

Example-based speech enhancement is a promising approach for coping with highly non-stationary noise. Given a noisy speech input, it first searches in noisy speech corpora 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. This framework works well if the noisy speech corpora contain the noise included in the input. However, it is impossible to prepare corpora that cover all types of noisy environments. Moreover, the example search is usually performed using noise sensitive mel-frequency cepstral coefficient features (MFCCs). Consequently, a mismatch between an input and the corpora is inevitable. This paper proposes using bottleneck features (BNFs) extracted from a deep neural network (DNN) acoustic model for the example search. Since BNFs have good noise robustness (invariance), the mismatch is mitigated and thus a more accurate example search can be performed. Experimental results on the Aurora4 corpus show that the example-based approach using BNFs greatly improves the enhanced speech quality compared with that using MFCCs. It also consistently outperforms a conventional DNN-based approach, i.e. a denoising autoencoder.