A Deep Neural Network Approach for Missing-Data Mask Estimation on Dual-Microphone Smartphones: Application to Noise-Robust Speech Recognition

The inclusion of two or more microphones in smartphones is becoming quite common. These were originally intended to perform noise reduction and few benefit is still being taken from this feature for noise-robust automatic speech recognition (ASR). In this paper we propose a novel system to estimate missing-data masks for robust ASR on dual-microphone smartphones. This novel system is based on deep neural networks (DNNs), which have proven to be a powerful tool in the field of ASR in different ways. To assess the performance of the proposed technique, spectral reconstruction experiments are carried out on a dual-channel database derived from Aurora-2. Our results demonstrate that the DNN is better able to exploit the dual-channel information and yields an improvement on word accuracy of more than 6% over state-of-the-art single-channel mask estimation techniques.