Speech Enhancement Using EMD in Low SNR Environment

Traditional spectral subtraction (SS) performs badly in low SNR environment. In this paper, we discuss the application of Empirical Mode Decomposition (EMD) in speech enhancement, and then introduce two approaches to speech enhancement using EMD. Based on EMD, the noise reduction scheme is fully data-driven. The basic principle of the method is the reconstruction of signal with the intrinsic mode functions, which are used SS or multi-band SS each. The improvements are more accurate noise estimation by using voice activity detection (VAD). It makes denoising more accurate and reasonable. Comparisons between four methods are performed under white Gaussian noise which is from Noisex92 database. The results demonstrate that the introduced methods achieve better speech quality or intelligibility in low SNR environment.