Speech synthesis based on feature extraction to enhance noise-corrupted speech

Presents a promising new alternative for improving the intelligibility of noise-corrupted speech. The technique relies on the extraction of speech features known to be important to speech intelligibility. These features are translated into parameters suitable to drive a formant-based speech synthesizer. The basic philosophy of the approach is that clean speech can be synthesized from features extracted from noisy speech providing the extracted features are sufficiently accurate. The features extracted are the fundamental frequency, voiced/unvoiced decision, and formant frequencies. Preliminary results demonstrate that intelligible speech can be synthesized from the information provided by the feature extraction algorithms, and that the proposed speech enhancement method has the potential to improve the intelligibility of speech corrupted by noise