HMM-based speech enhancement using pitch period information in voiced speech segments

The HMM-based scheme uses hidden Markov models (HMM) to control a state-dependent Wiener filter, which is used to process the noisy speech signal. This scheme gives enhanced speech signals without the annoying tonal artefacts ('musical noise') of the spectral subtraction approach. However, parts of the enhanced signal often sound rough or hoarse. In this paper, it is shown that this effect occurs because the noise between the harmonics of voiced speech segments is not removed by the Wiener filter. An algorithm is proposed, which uses pitch period information, and which is based on least squares (LS) estimation, to remove these noise components. Moreover, it is shown that the estimation involving low-energy states of the speech HMM is not reliable, and therefore a noise floor is inserted during low-energy speech segments instead of filtering the signal.