Sixth European Conference on Speech Communication and Technology
In speech recognition, acoustic modeling mismatches for the shapes of distributions are unavoidable when we use fixed-shape base distributions, such as the Gaussian. This paper proposes a new base distribution called the "generalized Laplacian distribution (GLD)", which offers more flexibility in terms of the distribution shape. The effectiveness of the GLD-based modeling is proven through speech recognition experiments.
Full Paper (PDF)
Bibliographic reference. Nakamura, Atsushi / Matsui, Tomoko (1999): "Acoustic modeling based on a generalized laplacian distribution", In EUROSPEECH'99, 1347-1350.