A Priori SNR Estimation Using an Artificial Neural Network

Inhalt:
Apart from noise power spectral density estimation and spectral weighting rule computation, the performance of a speech enhancement system is also highly dependent on the a priori signal-to-noise ratio (SNR) estimation. In this contribution, we present a data-driven approach to estimate the a priori SNR. Two trained artificial neural networks are employed in the proposed algorithm, one under hypothesis of speech presence, and one under hypothesis of speech absence. The neural networks use both additive components of the classical decision-directed a priori SNR estimator by Ephraim and Malah as the input signals to deliver new a priori SNR estimates at their output nodes. Being incorporated with a wide range of weighting rules, e.g., the minimum mean square error (log) spectral amplitude estimator, Wiener filter, or the super Gaussian joint maximum a posteriori estimator, the combination of the new SNR estimates in speech presence and absence reduces speech distortion, particularly in speech onset, while maintaining a high level of noise attenuation in speech absence.