Prediction-based time-varying filtering for speech enhancement in cellular telephony

The excitation signal contained in a speech signal and the vocal tract parameters determining the spectral envelope can be extracted by linear prediction (LP). In order to enhance noisy speech signals, time-varying filtering, based on the S-method as a time-frequency analysis tool, is applied to the excitation signal and to the spectral envelope in parallel. This approach is tested with noisy speech signals apparent in hands-free telephone systems.