On the optimization of sigmoid function for speech enhancement

This paper develops a methodology to optimize sigmoid function parameters based on a weighted sum of two objective measures, which are the perceptual evaluation of speech quality (PESQ) measure and the log-likelihood ratio (LLR) measure. The sigmoid function has been investigated for speech enhancement as an alternative gain function to the conventional MMSE function and the spectral subtraction function. The benefit of using this function is that it has tunable parameters for both its slope and its mean. It also provides a potential to preserve more speech signal at high SNR level. The SNR estimate and the gain function impact the value of the objective measures such as PESQ and LLR, and provide varying subjective quality. Thus, by studying the relationship between the SNR estimate and the gain function, the performance of a single channel speech enhancement scheme can be optimized. Here, we aim to optimize the parameters of sigmoid function for different types of noise conditions and SNRs. Subjective listening tests demonstrate a significant improvement in the objective measures with proper choice of parameters.