Bounded conditional mean imputation with an approximate posterior

Missing-feature imputation or reconstruction is used in noiserobust automatic speech recognition to recover the unobserved clean speech information. Reconstruction methods often use the noise-corrupted observations and a clean speech prior to calculate a point estimate for the unobserved clean speech features, whereas the approach proposed in this work associates the unobserved clean speech features with a full posterior distribution. The posterior mean can be used as a clean speech estimate in bounded conditional mean imputation and the posterior variance can be included as observation uncertainties. The proposed method is evaluated in a large-vocabulary noise-robust speech recognition task with speech data recorded in real noisy environments.