A great number of methods for multichannel audio source separation are based on probabilistic approaches in which the sources are modeled as latent random variables in a time-frequency (TF) domain. For reverberant mixtures, most of the methods approximate the time-domain convolutive mixing process in the TF-domain, assuming short mixing filters. The TF latent sources are then inferred from the TF mixture observations. In this paper we propose to infer latent TF sources from the time-domain observations. This approach allows us to exactly model the convolutive mixing process. The inference procedure rely on a variational expectation-maximization algorithm. In significant reverberation conditions, we show that our approach leads a Signal-to-Distortion Ratio improvement of 5.5 dB.