Single-Channel Noise Reduction with Short Delay

The objective of this thesis was to develop a single channel noise reduction method, capable of handling the short delay constraints of a digital hearing aid. Two candidates were developed and evaluated, both based on a previously proposed long-delay algorithm called Perceptually Optimised Spectral Subtraction, or shortly, PSS. The first method, PSSLP (PSS with LowPass filtering of filter coefficients), shortened the delay and compensated for the heavily fluctuating adaptive filter using a smoothing on the noise reduction filter coefficients. The second method, PSSW (PSS using frequency Warping), utilized a frequency warping technique in order to adapt the frequency analysis of the algorithm to the non-uniform frequency analysis of the human auditory system. The PSSW-method also reduced the signal delay to acceptable levels. Using both objective and informal subjective evaluations, and comparing different algorithm properties, the PSSW-method was found to provide the highest noise reduction of the two methods. In low-frequency noise environments the PSSW-method performed well and did not distort the signal. In other environments the PSSW-method introduced disturbing signal distortion and left unnatural-sounding residual noise. The PSSLPmethod provided a more robust performance for all environments. The method resulted in an apparent noise reduction and preserved a natural representation of both speech and residual noise. The PSSLP-method is thus proposed for more generic single channel noise reduction in short-delay applications.