Denoising Speech Signals for Digital Hearing Aids: A Wavelet Based Approach
Nathaniel A. Whitmal, Janet C. Rutledge, Jonathan Cohen · Applied and numerical harmonic analysis · 2011
This study describes research developing a wavelet based, single microphone noise reduction algorithm for use in digital hearing aids. The approach reduces noise by expanding the observed speech in a series of implicitly filtered, shift-invariant wavelet packet basis vectors. The implicit filtering operation allows the method to reduce correlated noise while retaining low-level high-frequency spectral components that are necessary for intelligible speech. Recordings of speech in automobile road noise at signal-to-noise ratios of 0, 5, 10, 15, and 20 dB were used to evaluate the new method. Objective measurements indicate that the new method provides better noise reduction and lower signal distortion than previous wavelet-based methods, and produces output free from audible artifacts of conventional FFT methods. However, trials of the Revised Speech Perception in Noise test with the new algorithm showed no significant improvement in speech perception. Subsequent analysis has shown that the algorithm imposes physical attenuation on low-intensity components that mimics the perceptual effects of mild hearing loss. These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.