Multiple fixed beamformers with a spacial Wiener-form postfilter for far-field speech recognition

Far-field speech recognition is becoming a hot topic in research and industrial applications. In this paper, in order to improve far-field speech recognition performance, we propose to use multiple fixed beamformers with a spacial Wiener-form postfilter (MFB-SWP) to suppress noise and interference. Our proposed method consists of two parts, beamforming and post-filter estimation. First, multiple fixed beamformers are designed and each of them aims at one specific direction. Next the target speech is estimated using the fixed beamformer aiming to the target direction, and the noise and interference signals are estimated using the remaining beamformers. After that, we calculate a spacial Wiener-form time-frequency and frame-level gains, as postfilter to further reduce the residual noise and interference. Compared with a single fixed beamformer, the proposed MFB-SWP method can suppress noise and interference significantly. It is also computationally more efficient comparing with other adaptive beamforming methods. Our experiments showed that proposed method achieved 16-50% relative character error rate (CER) reduction compared with using the single fixed beamformer only.