International Journal of Science, Engineering and TechnologyAuthors: Kapil Dev Tyagi
Abstract: Single-channel speech enhancement remains a core problem in hands-free communication, hearing assistance, and automatic speech recognition front-ends, where only one corrupted observation is available. Classical spectral subtraction attains high segmental signal-to-noise ratio but injects annoying "musical noise", whereas decision-directed Wiener filtering smooths the spectrum at the cost of speech attenuation. This paper proposes Adaptive Perceptual Spectral Subtraction (APSS), a short-time spectral gain that unifies four ideas: a decision-directed a priori signal-to-noise ratio driving a parametric square-root Wiener gain that preserves speech, a soft speech-presence indicator derived from the a posteriori SNR, a masking-guided suppression stage that removes only inaudible residual energy, and an SNR-adaptive temporal smoothing of the gain that erases isolated spectral peaks without smearing onsets. The method is evaluated on controlled synthetic speech corrupted by white, pink and babble noise from -5 to 15 dB. On stationary noise APSS attains the highest segmental SNR (7.9 dB, versus 7.8 dB for spectral subtraction and 4.5 dB for Wiener) while reducing log-spectral distortion by roughly 9% relative to spectral subtraction, and it keeps the musical-noise kurtosis ratio about seven times lower than the Wiener baseline. The results show that APSS occupies a favourable point on the distortion-versus-musical-noise trade-off surface that neither baseline reaches.
DOI: http://doi.org/10.5281/zenodo.20838131
