On generating mixing noise signals with basis functions for simulating noisy speech and learning dnn-based speech enhancement models

Abstract

We first examine the generalization issue with the noise samples used in training nonlinear mapping functions between noisy and clean speech features for deep neural network (DNN) based speech enhancement. Then an empirical proof is established to explain why the DNN-based approach has a good noise generalization capability provided that a large collection of noise types are included in generating diverse noisy speech samples for training. It is shown that an arbitrary noise signal segment can be well represented by a linear combination of microstructure noise bases. Accordingly, we propose to generate these mixing noise signals by designing a set of compact and analytic noise bases without using any realistic noise types. The experiments demonstrate that this noise generation scheme can yield comparable performance to that using 50 real noise types. Furthermore, by supplementing the collected noise types with the synthesized noise bases, we observe remarkable performance improvements implying that not only a large collection of real-world noise signals can be alleviated, but also a good noise generalization capability can be achieved.