Contribution
Evaluation of Short-Term Features for Online Estimation of the A Priori Probability for Speech Absence in Multichannel Speech Presence Probability Estimators
* Presenting author
Abstract:
Speech presence probability (SPP) of the target speech in a mixture of signals is of crucial importance in many speech communication applications such as noise reduction and speech enhancement. Based on the Gaussian model multichannel speech presence probability the aim of this paper is to compensate for the fact that this algorithm only achieves a minimum SPP equal to the counterprobability of the a priori probability of speech absence. This leads to a distorted SPP when no speech is present. Instead of a fixed a priori probability of speech absence it is proposed to estimate the a priori probability online with short-term features including short-term energy and a spectral flatness measure as used in many voice activity detection algorithms. This leads to a more realistic SPP estimation when speech is absent while only introducing a negligible increase in computational complexity.