Evaluation of noise reduction techniques for digital hearing aids

Abstract

Individuals with sensorineural hearing loss have increased difficulty in understanding speech in noisy backgrounds. To combat this issue, there has been a major thrust in recent years toward the development of noise reduction algorithms. The goals of this paper are to quantify the relative benefits of different single-microphone noise reduction algorithms, and to investigate the interaction between the noise reduction and dynamic range compression algorithms. Noise reduction techniques evaluated in this paper include spectral subtraction-based techniques, a wavelet-packet-based technique and a matching pursuit-based technique. All algorithms were tested with HINT signals with SNR levels ranging from −5 to 15 dB, and two different noise types viz. the speech-shaped noise and multi-talker babble. Performance was quantified using the ITU standardized PESQ measure which computes the perceptual similarity between the enhanced signal and the original signal. Initial PESQ results showed that the spectral subtraction-based techniques perform superior to that of the wavelet-packet and matching pursuit-based approaches and that the compression time constants have an impact on the overall performance. Perceptual data collected from hearing impaired listeners on sound quality and noise reduction performance will be presented and their correlation with the objective measurements will be discussed.