Publication Details
Using Smoothed Heteroscedastic Linear Discriminant Analysis in Large Vocabulary Continuous Speech Recognition System
Burget Lukáš, doc. Ing., Ph.D. (DCGM FIT BUT)
Černocký Jan, prof. Dr. Ing. (DCGM FIT BUT)
speech recognition, LVCSR, HLDA, feature transform, dimensionality reduction
In this work, we verify that SHLDA can be advantageously used also for Large Vocabulary
Continuous Speech Recognition.
In the state-of-the-art speech recognition systems, Heteroscedastic Linear Discriminant Analysis (HLDA)
is becoming popular technique allowing for feature decorrelation and dimensionality reduction.
However, HLDA relies on statistics, which may not be reliably estimated when only limited amount of
training data is available. Recently, Smoothed HLDA (SHLDA) was proposed
as a robust modification of
HLDA. Previously, SHLDA was successfully used for feature combination in
small vocabulary recognition
experiments. In this work, we verify that SHLDA can be advantageously used also for Large Vocabulary
Continuous Speech Recognition.
@INPROCEEDINGS{FITPUB7929,
author = "Martin Karafi\'{a}t and Luk\'{a}\v{s} Burget and Jan \v{C}ernock\'{y}",
title = "Using Smoothed Heteroscedastic Linear Discriminant Analysis in Large Vocabulary Continuous Speech Recognition System",
pages = 8,
booktitle = "2nd Joint Workshop on Multimodal Interaction and Related Machine Learning Algorithms",
series = "tento \v{c}l\'{a}nek nebyl za\v{r}azen mezi Revised Selected Papers, nevy\v{s}el v LNCS 3869",
year = 2005,
location = "Edinbourgh, Scotland, GB",
publisher = "University of Edinburgh",
language = "english",
url = "https://www.fit.vut.cz/research/publication/7929"
}