End-to-End Probabilistic Inference for Nonstationary Audio Analysis