Comparison of predictive measures of speech recognition after noise reduction processing

Area of Science:

• Audiology
• Speech processing
• Acoustics

Background:

• Single-channel noise reduction (NR) aims to improve speech recognition in noisy environments.
• Predicting the effectiveness of NR algorithms is crucial for hearing aid development.
• Existing prediction measures vary in their accuracy.

Purpose of the Study:

• To evaluate various measures for predicting the speech-recognition benefit of single-channel NR.
• To compare predicted benefits with experimentally measured outcomes.
• To identify the most accurate predictive measures for NR algorithms.

Main Methods:

• Three NR algorithms and a reference condition were tested.
• Twenty listeners with impaired hearing and ten with normal hearing participated.
• An adaptive speech test determined signal-to-noise ratios for 80% speech recognition accuracy.

Main Results:

• Nine predictive measures, including two variants, were evaluated against experimental data.
• The Coherence Speech Intelligibility Index (CSII) provided the most accurate predictions.
• Measures based on speech-noise correlation and short-time analysis showed promise.

Conclusions:

• The CSII is a promising measure for predicting speech recognition benefits from NR.
• Measures analyzing short-time speech and noise characteristics are effective predictors.
• Accurate prediction of NR benefits can guide the development of improved hearing technologies.