Publication Details
Region Dependent Linear Transforms in Multilingual Speech Recognition
Janda Miloš, Ing.
Černocký Jan, prof. Dr. Ing. (DCGM)
Burget Lukáš, doc. Ing., Ph.D. (DCGM)
HLDA, Region Dependent Transforms, MinimumPhone Error, fMPE, multilingual speech recognition
In today's speech recognition systems, linear or nonlinear transformationsare usually applied to post-process speech features forminginput to HMM based acoustic models. In this work, we experimentwith three popular transforms: HLDA,MPE-HLDA and Region DependentLinear Transforms (RDLT), which are trained jointly withthe acoustic model to extract maximum of the discriminative informationfrom the raw features and to represent it in a form suitablefor the following GMM-HMM based acoustic model. We focus onmulti-lingual environments, where limited resources are availablefor training recognizers of many languages. Using data from GlobalPhonedatabase, we show that, under such restrictive conditions,the feature transformations can be advantageously shared across languagesand robustly trained using data from several languages.
@inproceedings{BUT91480,
author="Martin {Karafiát} and Miloš {Janda} and Jan {Černocký} and Lukáš {Burget}",
title="Region Dependent Linear Transforms in Multilingual Speech Recognition",
booktitle="Proc. International Conference on Acoustics, Speech, and Signal Processing 2012",
year="2012",
pages="4885--4888",
publisher="IEEE Signal Processing Society",
address="Kyoto",
doi="10.1109/ICASSP.2012.6289014",
isbn="978-1-4673-0044-5",
url="http://www.fit.vutbr.cz/research/groups/speech/publi/2012/karafiat_icassp2012_0004885.pdf"
}