Perceptually motivated pre-filter for speech enhancement using Kalman filtering

This paper proposes a novel pre-filter for Kalman filter based speech enhancement. Our aim is to reduce coloured noise, while retaining speech quality by exploiting the properties of the human auditory system. The proposed pre-filter uses temporal and simultaneous masking thresholds to shape the noisy speech spectrum in order to obtain a better estimate of the AR coefficients. These coefficients are then used to obtain a perceptual filter to weight the noise spectrum before applying the Kalman filter. The proposed noise reduction technique is compared to the Wiener filter, Perceptual Wiener filter and the standard Kalman Filter in 5 dB car noise environments. PESQ scores and subjective test results show that the proposed pre-filter based enhancement outperforms other common techniques.