Fast joint DOA and pitch estimation using a broadband MVDR beamformer

Abstract

The harmonic model, i.e., a sum of sinusoids having frequencies that are integer multiples of the pitch, has been widely used for modeling of voiced speech. In microphone arrays, the direction-of-arrival (DOA) adds an additional parameter that can help in obtaining a robust procedure for tracking non-stationary speech signals in noisy conditions. In this paper, a joint DOA and pitch estimation (JDPE) method is proposed. The method is based on the minimum variance distortionless response (MVDR) beamformer in the frequency-domain and is much faster than previous joint methods, as it only requires the computation of the optimal filters once per segment. To exploit that both pitch and DOA evolve piece-wise smoothly over time, we also extend a dynamic programming approach to joint smoothing of both parameters. Simulations show the proposed method is much more robust than parallel and cascaded methods combining existing DOA and pitch estimators.