Matrix sentence intelligibility prediction using an automatic speech recognition system.

Area of Science:

• Speech Processing
• Acoustics
• Computational Linguistics

Background:

• Predicting speech intelligibility in noise is crucial for audiology and assistive technologies.
• Traditional methods often rely on human listeners or simplified acoustic indices.
• Evaluating automated prediction systems is essential for advancing speech perception research.

Purpose of the Study:

• To assess the feasibility of using an automatic speech recognition (ASR) system to predict speech reception thresholds (SRTs) in the German matrix sentence test.
• To compare ASR-based predictions with empirical data and the speech intelligibility index (SII) across various stationary noise conditions.

Main Methods:

• An ASR system utilizing Mel-frequency cepstral coefficients and Hidden Markov models was developed.
• The ASR system was trained and tested on the German matrix sentence test with varying signal-to-noise ratios in seven distinct noise conditions.
• ASR predictions were compared against published empirical data from normal-hearing listeners and SII predictions.

Main Results:

• ASR-based SRT predictions demonstrated a high and significant correlation (R² = 0.95, p < 0.001) with empirical data.
• ASR predictions significantly outperformed SII predictions, which showed no correlation (R² = 0.00, p = 0.987) with empirical data.
• The ASR system accurately predicted performance across different stationary noise types.

Conclusions:

• An ordinary, reference-free ASR system can effectively predict SRTs for the German matrix test in stationary noise for normal-hearing listeners.
• The predictive accuracy relies on the acoustical properties of speech and noise signals, with minimal assumptions about human auditory processing.
• This ASR-based approach offers a promising, automated alternative for assessing speech intelligibility in challenging acoustic environments.