%Aigaion2 BibTeX export from Idiap Publications
%Thursday 21 November 2024 12:47:11 PM

@TECHREPORT{Luyet_Idiap-RR-05-2016,
         author = {Luyet, Gil},
       projects = {Idiap},
          month = {3},
          title = {Low-Rank Representation For Enhanced Deep Neural Network Acoustic Models},
           type = {Idiap-RR},
         number = {Idiap-RR-05-2016},
           year = {2016},
    institution = {Idiap},
       abstract = {Automatic speech recognition (ASR) is a fascinating area of research towards realizing humanmachine
interactions. After more than 30 years of exploitation of Gaussian Mixture Models (GMMs),
state-of-the-art systems currently rely on Deep Neural Network (DNN) to estimate class-conditional
posterior probabilities. The posterior probabilities are used for acoustic modeling in hidden Markov
models (HMM), and form a hybrid DNN-HMM which is now the leading edge approach to solve
ASR problems.
The present work builds upon the hypothesis that the optimal acoustic models are sparse
and lie on multiple low-rank probability subspaces. Hence, the main goal of this Master project
aimed at investigating different ways to restructure the DNN outputs using low-rank representation.
Exploiting a large number of training posterior vectors, the underlying low-dimensional subspace
can be identified, and low-rank decomposition enables separation of the “optimal” posteriors from
the spurious (unstructured) uncertainties at the DNN output.
Experiments demonstrate that low-rank representation can enhance posterior probability estimation,
and lead to higher ASR accuracy. The posteriors are grouped according to their subspace
similarities, and structured through low-rank decomposition. Furthermore, a novel hashing technique
is proposed exploiting the low-rank property of posterior subspaces that enables fast search
in the space of posterior exemplars.},
            pdf = {https://publications.idiap.ch/attachments/reports/2016/Luyet_Idiap-RR-05-2016.pdf}
}