Evaluation of the sparse coding shrinkage noise reduction algorithm in normal hearing and hearing impaired listeners

Area of Science:

• Signal processing
• Auditory perception
• Acoustics

Background:

• Many single-channel noise reduction methods enhance speech quality but not intelligibility, especially with similar speech and noise spectra.
• Existing intelligibility-enhancing methods often require prior statistical knowledge, limiting real-world application.
• Hearing-impaired (HI) listeners experience greater speech intelligibility deficits in noise compared to normal-hearing (NH) listeners.

Purpose of the Study:

• To develop and evaluate a model-based single-channel noise reduction algorithm, sparse coding shrinkage (SCS), for improving speech intelligibility in noisy conditions.
• To compare the SCS algorithm's performance against a state-of-the-art Wiener filtering approach using speech intelligibility tests.
• To investigate the effectiveness of these algorithms for both NH and HI listeners across different noise types and signal-to-noise ratios (SNRs).

Main Methods:

• The study employed a model-based sparse coding shrinkage (SCS) algorithm that utilizes statistical signal information without requiring prior knowledge.
• Speech intelligibility tests were conducted with both NH and HI listeners comparing the SCS algorithm against Wiener filtering.
• Performance was evaluated in stationary and fluctuating (babble) noise conditions at various input signal-to-noise ratios (SNRs).

Main Results:

• The SCS algorithm demonstrated improved speech intelligibility in stationary noise, performing comparably to the Wiener filtering algorithm.
• Both SCS and Wiener filtering improved intelligibility for HI listeners but not significantly for NH listeners.
• Performance gains were more pronounced in stationary noise than in fluctuating noise and were better at higher SNRs, benefiting HI listeners more than NH listeners.

Conclusions:

• The SCS algorithm is a promising alternative to Wiener filtering for speech intelligibility enhancement, particularly for HI listeners.
• HI listeners require distinct signal processing strategies compared to NH listeners, with performance influenced by input SNR rather than solely hearing loss level.
• The effectiveness of noise reduction algorithms may vary with the degree of hearing loss, suggesting potential for greater benefit in individuals with more severe hearing impairments.