Evaluation of the sparse coding shrinkage noise reduction algorithm for the hearing impaired

Although there are numerous single-channel noise reduction strategies to improve speech perception in a noisy environment, most of them can only improve speech quality but not improve speech intelligibility for normal hearing (NH) or hearing impaired (HI) listeners. Exceptions that can improve speech intelligibility currently are only those that require a priori statistics of speech or noise. Most of the noise reduction algorithms in hearing aids are adopted directly from the algorithms for NH listeners without taking into account of the hearing loss factors within HI listeners. HI listeners suffer more in speech intelligibility than NH listeners in the same noisy environment. Further study of monaural noise reduction algorithms for HI listeners is required. The motivation is to adapt a model-based approach in contrast to the conventional Wiener filtering approach. The model-based algorithm called sparse coding shrinkage (SCS) was proposed to extract key speech information from noisy speech. The SCS algorithm was evaluated by comparison with another state-of-the-art Wiener filtering approach through speech intelligibility and quality tests using 9 NH and 9 HI listeners. The SCS algorithm matched the performance of the Wiener filtering algorithm in speech intelligibility and speech quality. Both algorithms showed some intelligibility improvements for HI listeners but not at all for NH listeners. The algorithms improved speech quality for both HI and NH listeners. Additionally, a physiologically-inspired hearing loss simulation (HLS) model was developed to characterize hearing loss factors and simulate hearing loss consequences. A methodology was proposed to evaluate signal processing strategies for HI listeners with the proposed HLS model and NH subjects. The corresponding experiment was performed by asking NH subjects to listen to unprocessed/enhanced speech with the HLS model. Some of the effects of the algorithms seen in HI listeners are reproduced, at least qualitatively, by using the HLS model with NH listeners. Conclusions: The model-based algorithm SCS is promising for improving performance in stationary noise although no clear difference was seen in the performance of SCS and a competitive Wiener filtering algorithm. Fluctuating noise is more difficult to reduce compared to stationary noise. Noise reduction algorithms may perform better at higher input signal-to-noise ratios (SNRs) where HI listeners can get benefit but where NH listeners already reach ceiling performance. The proposed HLS model can save time and cost when evaluating noise reduction algorithms for HI listeners.EThOS - Electronic Theses Online ServiceGBUnited Kingdo