Integrated acoustic echo and background noise suppression based on stacked deep neural networks

초록

In this paper, a regression-based integrated acoustic echo and background noise suppression algorithm was proposed through the use of a deep neural network (DNN) with a multi-layer deep architecture. Motivated by an idea that DNNs are a superior hierarchical generative model for modeling the complex relationships between input features and desired target features through its multiple nonlinear hidden layers, a stacked DNN is developed in a sequential fashion such that the DNN for noise suppression is followed by the DNN for acoustic echo suppression. This algorithm is compared to a single DNN-based integrated system to simultaneously suppress acoustic echoes and noise. When developing the DNN-based regression technique using our approach, spectral envelop estimation is a crucial point for which log-power spectra (LPS) are used as features in order to determine the gain, which ensured nonlinear mapping from the LPS of the frames contaminated by echoes and noise to the LPS of the echo- and noise-free frames. This leads to the successful reduction of acoustic echoes and background noise without an additional double-talk detection algorithm. Additionally, an augmented feature technique is adopted to use additional knowledge derived from conventional noise and acoustic echo suppression techniques when designing the DNN architecture in our algorithm. The proposed DNN-based integrated system to suppress acoustic echoes and noise was evaluated in terms of objective measures and demonstrated a significant improvement over conventional integrated algorithms.

키워드