Enhancing the Noise Immunity in Speech Signal by Using Combined Filtering Technique

Abstract

            Improving the noise reduction and improving voice signal efficiency in speech processing is one of the most significant considerations in today's studies. The key to real-time signal processing is the reduction in speech signals by noise. Due to the complexity of the identification of speakers and the specific feature of a public channel, noise decrease in the voice signal processing community has always been a challenge. To solve this problem, we proposed a new technique of noise reduction for speech signalling established on complete empirical ensemble mode decomposition with adaptive noise (CEEMDAN) and least mean square adaptive filter (LMSAF). This technique of sound reductions, known as CEEMDAN -LMSAF, has three main advantages, including enhanced algorithms for the decomposition of the empirical mode (EMD) and the EMD ensemble (EEMD), which can reduce mixing and prevent the loss of vibrational mode decomposition (VMD). LMSAF also overcomes the range of the decomposition amount and the basis for reducing wavelet noise with regard to noise reduction in bright IMFs. First, the original signal is broken down into IMFs, which can be distributed into bright IMFs and actual IMFs. LMSAF is then used to sense noisy IMFs and to remove modules of noisy IMFs. Finally, there is a reconstruction of both denounced noisy IMFs and real IMFs, and the last signal denounced. Through comparing simulation signals and noise signals, which have a greater impact on noises reduction and a convenient implementation, the validity of CEEMDAN-LMSAF can be shown relative to other noise reduction approaches. CEEMDAN-LMSAF also provides the accurate basis from which speech signals are detected, extracted, categorised and recognised.