Improving speech intelligibility in background noise by SII-dependent amplification and compression

In many speech communication applications it is of great interest to achieve a high intelligibility to ensure good communication. However, in these applications speech is often disturbed by additive noise and/or reverberation. Therefore, it is desirable to develop algorithms that are able to maintain a high intelligibility in such disturbed scenarios. While amplifying the speech to achieve good signal-to-noise ratios (SNR) is an easy approach, it is often not applicable due to technical limitations of the amplification system or unpleasantly high sound levels. Consequently, algorithms that increase speech intelligibility while maintaining equal powers are preferable. Several algorithms have been proposed in the past that use either frequency-dependent amplification, dynamic range compression, transient amplification, or modulations filtering techniques. The first attempt to investigate the effect of different signal processing strategies on speech intelligibility was made by Licklider and Pollack [1]. While they did not consider any additive noise or reverberation they could demonstrate that speech intelligibility in quiet is not necessarily affected by strategies such as high-pass or lowpass filtering and clipping. Niederjohn and Grotelueschen [2] proposed a preprocessing algorithm that uses high-pass filtering followed by static rapid amplitude compression. They observed an increase in speech intelligibility for preprocessed speech in white noise over the unprocessed speech at the same SNR. Zorila et al. [4] adopted the idea of dynamic range compression as a mean to increase the speech intelligibility. They used a static input-output characteristic for their dynamic range compression and used several frequency-dependent amplification steps prior to compression. Recently, Sauert and Vary proposed an algorithm that uses timeand frequency-dependent amplification of the speech signal aiming to maximize the SII [3]. However, this approach suffers from spectral adaption to the background noise. Therefore in a recent approach they considered an SNR-dependent transition between SIIweighted and unity-weighting of the speech signal [6]. In this contribution we describe an algorithm and its