Probabilistic spectral gain modification applied to beamformer-based noise reduction in a car environment

In this paper, we propose a new noise reduction technique for improving the performance of conventional beamformers used in cars. To this end, the probabilistic discrimination on target-directional signal presence is used as a parameter for the spectral gain modification (SGM) of a beamformer output. The direction of arrival (DOA)-based target-to-non-target-directional signal ratio (TNR) is first estimated by using spatial cues such as phase differences from multiple microphone signals. Next, the estimated TNR is utilized to estimate the target-directional signal presence probabilities (TDSPPs) that include global and local terms. The performance of the proposed SGM is evaluated by the degree of noise reduction, average and segmental signal-to-noise ratio (SNR), as well as perceptual evaluation of speech quality (PESQ) scores under car noise conditions whose SNR varies from -5 to 20 dB. As a result, it is shown that the proposed SGM significantly improves the target-directional signal enhancing performance against conventional beamformers, i.e., delay-and-sum beamformer (DSB), super directive beamformer (SDB) and generalized sidelobe canceller (GSC), for all input SNRs.