Output details
11 - Computer Science and Informatics
University of Edinburgh
Joint Uncertainty Decoding for Noise Robust Subspace Gaussian Mixture Models
<22> Originality: This work develops and evaluates model-based noise compensation for subspace GMM acoustic models which have been demonstrated to extend the state of the art in speech recognition. The originality lies in the computationally feasible approach to JUD-based noise compensation.
Significance: Robustness to noise is the biggest challenge in automatic speech recognition. This work solves some key technical problems in applying model-based noise compensation to subspace GMMs.
Rigour: State-of-the-art experimental results when evaluated on the standard Aurora-4 corpus. Provides detailed derivations for JUD transform estimation with subspace GMMs.