Improving speech intelligibility in noise by SII-dependent preprocessing using frequency-dependent amplification and dynamic range compression

Abstract

In this contribution, a new preprocessing algorithm to improve speech intelligibility in noise is proposed, which maintains the signal power before and after processing. The proposed AdaptDRC algorithm consists of two timeand frequency-dependent stages, which are both functions of the estimated SII. The first stage applies a timeand frequency-dependent amplification, while the second stage applies a timeand frequency-dependent dynamic range compression (DRC). Experiments with a competing speaker (CS) and a speech-shaped noise (SSN) show an increase in speech intelligibility for a wide range of SNRs for four different objective measures that are correlated with speech intelligibility. Listening tests conducted within the framework of the Hurricane Challenge with 175 subjects confirm these findings and show improvements of up to 20.5% in intelligibility for SSN and 12.3% for CS.