Noise power spectral density estimation for binaural noise reduction exploiting direction of arrival estimates
Daniel Marquardt, Simon Doclo · 2017
Noise reduction algorithms for head-mounted assistive listening devices are crucial to improve speech quality and intelligibility in background noise. For binaural hearing devices with one microphone per device, the noise power spectral density (PSD) is commonly estimated using various assumptions about the acoustic scenario. Since these methods lack robustness if the underlying assumptions are not satisfied, alternatively the noise PSD can be estimated at the output of a blocking matrix, however requiring an estimate of the relative transfer function (RTF) or direction of arrival (DOA) of the desired speech source. For constructing the blocking matrix, in this paper we exploit RTF estimates using the covariance whitening method and DOA estimates obtained from a binaural DOA estimator using anechoic prototype acoustic transfer functions (ATFs). Simulation results in a realistic cafeteria scenario show that exploiting DOA estimates for binaural noise PSD estimation leads to an improved noise reduction performance, especially in the presence of directional interfering speakers.