Uncertainty decoding for DNN-HMM hybrid systems based on numerical sampling

In this article, we propose an uncertainty decoding scheme for DNN-HMM hybrid systems based on numerical sampling. A finite set of samples is drawn from the estimated probability distribution of the acoustic features and subsequently passed through feature transformations/extensions and the deep neural network (DNN). Then, the nonlinearly-transformed feature samples are averaged at the output of the DNN in order to approximate the posterior distribution of the context-dependent Hidden Markov Model (HMM) states. This concept is experimentally verified for the REVERB challenge task using a reverberation-robust DNN-HMM hybrid system: The numerical sampling is performed in the logmelspec domain, where we estimate the posterior distribution of the acoustic features by combining coherence-based Wiener filtering and uncertainty propagation. The experimental results highlight the good performance of the proposed uncertainty decoding scheme with significantly increased recognition accuracy even for a small number of feature samples.