Noise reduction based on robust speech and non-speech detection in vehicular environments

Background/Objectives A variety kinds of noise reduction approaches have achieved different performance according to noise types and user environments. Therefore, a suitable noise reduction approach needs to be investigated in consideration of a specific user environment. Methods/Statistical Analysis Most noise reduction methods are affected by the correctness of speech and non-speech detection, because a primary process of the methods is to estimate noise components from non-speech regions. In this study, we propose an efficient noise reduction approach to be adopted in vehicular environments. The proposed approach is based on robust speech and non-speech detection approach based on variance of spectral energy in frequency bin. Findings To observe the property of vehicular noise signals sophisticatedly, we recorded a set of noise sounds inside a car while driving the car. Vehicular noise signals have a general property of stationary noise types, indicating slowly changing signal characteristics. The property is observed more significantly in frequency domain compared to time domain. This observation concludes that in vehicular environments, a method of reducing stationary noise signals is quite applicable. The proposed approach considers the general tendency of spectral energy that speech regions indicate higher spectral variance than non-speech regions do. Thus, the method can improve the performance of noise reduction. Once a non-speech region is detected according to our mechanism, the spectral energy is regarded as the energy of noise components. So, the estimated spectral energy is used for suppressing noise components in subsequent speech regions. Improvements/Applications We evaluated the efficiency of the proposed approach via spectral analysis. The proposed voice activity detection method demonstrated superior noise reduction performance compared to the conventional method in terms of SNR.