Improved perceptually inspired speech enhancement using a psychoacoustic model

A speech enhancement algorithm is described that uses a psychoacoustic model to estimate speech cues in the presence of low SNRs. The motivation of the proposed algorithm is to model the peripheral of human auditory system and derive a perceptually improved solution for noise suppression. In the model, a detector for speech saliency is derived to measure the degree of conspicuousness of "significant" speech signals with the presence of background noise. Three biological correlates are exploited, including a spectral saliency determining the speech cues by frequency sensitivity of human auditory system in cochlear, a phoneme saliency indicating the perception discrimination of phonemes and an audibility saliency illuminating psychoacoustic masking properties by the frequency-to-place transformation of basilar membrane. The detector generates frequency-based soft-decisions that are used in determining the presence of speech cues and controlling the parameters of speech enhancement to preserve interested components.