Non-linear soft-sounds enhancement for near-end speech intelligibility improvement

The objective of this research is to modify the clean speech in a way that it will be more intelligible when it is played in noisy environment without increasing global signal-to-noise ratio. A new near-end speech enhancement algorithm is derived in this contribution based on an extrapolation technique. In this method speech energy is transferred from high energy regions of the speech signal to low energy regions by considering soft-sounds/strong-voiced components classification decisions into account. Variable amplification gain is derived and applied to the classified speech components depending on their original energy levels. The proposed algorithm does not require any information about input noise characteristics for near-end speech enhancement problem. The derived algorithm is combined with baseline near-end speech enhancement method as a post processing block for testing overall performance. Significant intelligibility improvements are observed with the proposed method over unprocessed noisy speech and considerable improvements are observed with combined method over recent version of the baseline method.