Hybrid Noise Reduction And Enhancement of Audio Quality using Deep Learning

Abstract

Good voice communication has become a top requirement in today’s rapidly growing world. Noise from the surroundings affects the quality of voice and audio signals in acoustic applications. The original voice signal that was broadcast can occasionally no longer be recovered. Acoustic Noise Cancellation (ANC) is a method for improving the quality of speech and audio signals by removing noise in the voice signal. The adaptive filter, a crucial component of the ANC, reduces noise without first knowing the difference between the signal and the noise. Conventional filters would cause the desired voice stream to be distorted. So, when speech and noise signals are random, adaptive filters are appropriate. A primary input with a damaged signal and a reference input with noise that is unknowably associated with the noise in the primary input make up the ANC’s two inputs. The reference input is adaptively filtered and subtracted from the primary input to get the clean speech signal estimation. The performance of the LMS and NLMS algorithms is significantly impacted by the step size and filter length M. Smaller mean square errors lead to longer convergence times (MSE). When it is large, the algorithm diverges, which reduces the adaptive filter’s efficacy. The trade-off between convergence time and MSE must thus be balanced when deciding on a step size, which is a difficult problem. Another practical difficulty is presented by selecting the filter’s tap length M. As filter length M rises, so do the filter’s convergence time and MSE. Consequently, a shorter-length filter is required. Unfortunately, deciding on the number of filter taps is mostly a matter of experience and trial and error. In LMS and NLMS algorithms, it can be challenging to choose the step size of the algorithm and the length of the adaptive filter M to provide the most excellent possible noise cancellation.