Binaural Deep Neural Network for Noise Robust Automatic Speech Recognition

Robust automatic speech recognition (ASR) is a challenge task, especially in noisy environments. The difference between the clean training speech model and the noisy speech model is a main factor to reduce the performance of ASR systems. The goal of a robust ASR system is getting the target speech energy distribution, which provides the discriminate information for the acoustic model. We use a binaural deep neural network (DNN) to estimate the energy of the target speech in the mixture through SNR estimation. Then the estimated target speech is used as the input of a convenient ASR system to improve the recognition accuracy. We use the ideal parameter mask as the DNN training goal, and cross entropy as the training cost function. Experiments show the robust ASR performance of the proposed algorithm with various signal to noise ratio conditions.