Speech enhancement based on auditory masking properties and log-spectral distance

This paper studies on auditory masking properties and redefines log-spectral distance. Combined with the minimum mean-square error log-spectral amplitude estimator, a speech enhancement algorithm using the auditory masking properties and the log-spectral distance is proposed. This algorithm capitalizes the log-spectral distance to detect voice activity, and updates the noise estimator in real time according to the presence or absence of speech. In speech section, the residual noise is suppressed by auditory masking properties. In non-speech section, the ratio between the log-spectral distance and the distance threshold is used to suppress the residual noise. The experimental results show that this algorithm can effectively suppress the residual noise, and protect the weak voice.