A Novel Method to Correct Steering Vectors in MVDR Beamformer for Noise Robust ASR

Abstract

Accurate steering vectors (SV) are key to many beamformers. However, reliable SV is not easy to obtain. In this work, we investigate a novel method to identify and correct phase errors in SV for MVDR beamforming. Our idea stems from the linear relationship in the phase of a microphone component in narrowband SVs across frequency, as modeled by acoustic transfer function. We utilize this property and feedforward neural nets to make phase prediction for the microphone components in SVs, and use the predicted phase selectively for phase error correction and MVDR beamforming. Our method is robust to large fluctuations in phase spectrum wrapped within [−π, π]. We have evaluated our approach on CHiME-3 and obtained improved performances on both word error rate and short-time objective intelligibility in low reverberant acoustic environments.