Abstract:
Clear speech recognition in a noisy environment can be challenging for people with hearing
impairment. In this thesis, noise reduction techniques have been investigated using the classical
approach in MATLAB to improve a digital hearing aid system. The first method focused on noise
reduction filters (amplitude, frequency, and de-noising) to reduce background noise, the second
approach solves the issue of single-microphone speech enhancement while the third method is using adaptive Least Mean Square (LMS) and Recursive Least Squares (RLS) algorithms for speech
enhancement.
Modern short-time noise reduction strategies are usually expressed as a spectral gain that is
proportional to the SNR. This problem is solved using the Two-step noise reduction (TSNR) technique, which maintains the Decision-Directed (DD) approach’s advantages. A second step refines
the calculation of the a priori SNR, eliminating the DD method’s bias and hence the reverberation
effect.Due to estimators’ unreliability for small signal-to-noise ratios, traditional short-time noise
reduction techniques, such as TSNR, introduce harmonic distortion in enhanced expression. This is
primarily due to the challenging task of estimating noise power spectrum density (PSD) in singlemicrophone schemes. The harmonic regeneration noise reduction (HRNR) method was investigated
and modified (HRNR) to solve this problem.
The simulation results show that the RLS algorithm demonstrates a significantly higher rate
of convergence of the weight coefficient to the optimal values compared to the LMS algorithm.
In order to achieve better results, different real-world noises used at different SNRs. The results
obtained show that the use of classical SNR approaches to improve speech enhancement provides
poor performance. In this thesis, studies of traditional noise reduction algorithms for digital hearing
aids were shown and their effectiveness was compared using SNR value.