Smoothing the acoustic spectral time series of speech signals for noise reduction

摘要

In this study, we propose a novel speech enhancement scenario that applies a low-pass filter to the spectrogram of a noise-corrupted signal along the temporal axis. According to our recently developed method, modulation-domain wavelet denoising (ModWD), the speech quality of noisy utterances can be promoted by alleviating the temporal-domain detail component of the respective spectrogram created by discrete wavelet transform (DWT), which implies the high (modulation) frequency part has little to do with speech quality and is relatively vulnerable to noise. For this reason, this study presents employing the conventional low-pass filter structure, including a moving-average (MV) filter and a median filter, to diminish the aforementioned high modulation frequency components in noise- corrupted spectrogram to see whether the speech quality can be improved. The preliminary experiments conducted on a subset of the Aurora-2 connected digit database show that the presented scenario can provide noisy signals with moderate improvement in speech quality, and the corresponding results are superior to those obtained by ModWD.