Improving speech synthesis for high intelligibility under adverse conditions

We investigate methods of improving the intelligibility of synthetic speech under noisy or low-fidelity acoustic conditions. Techniques explored improve speech in a natural manner, such that training won’t be required for the user to understand the enhanced speech. While the improvements are natural in this respect, the changes aren’t limited to creating only speech that is achievable by a human vocal tract. Modifications fall into three broad classes: increasing phoneme amplitude, altering spectral shape, and lengthening phoneme duration. Listening tests conducted in noisy and noise-free conditions demonstrate significant improvements to intelligibility for most of the subject phonemes.