Statistical model based SNR estimation method for speech signals

The performance of speech enhancement algorithms to a large extent is related to the employed signal-to-noise ratio (SNR) estimation techniques. Many of the existing SNR estimation techniques are based on approaches that require either an experimentally pre-specified weighting factor or prior assumptions of the parameters in the signal model. In this reported work, a closed form SNR estimator is derived by modelling the noisy speech signal as a generalised normal-Laplace distribution and estimating the variance of the signal and variance of the noise using high-order sample moments. The performance of the proposed technique is tested using real speech signals and compared with the well-known eigenvalue method.