|
Methods for combining language models in speech recognition
Simo Broman and Mikko Kurimo
In: Eurospeech 2005, 5-8 Sep 2005, Lisbon, Portugal.
Abstract Statistical language models have a vital part in contemporary speech
recognition systems, where its task is to score the word
hypothesis based on the linguistical knowledge. A lot of
language models have been presented in the literature. The
best results have been achieved when different language models
have been used together. Several combination methods have been
presented, but a thorough investigation of the methods has not
been done.
In this work, combination methods that have been used with
language models are studied. Also, a new approach based on
likelihood density function estimation using histograms is
presented. In addition to theoretical consideration, four
combining methods are evaluated in speech recognition experiments
and perplexity calculations that measure the quality of the
language models. The test data consist of Finnish news
articles. Four language models, that are presented in the work,
work as the component models.
In the perplexity experiments, all combining methods produced
statistically significant improvement compared to the 4-gram model
that worked as a baseline. The best result, 46 \% improvement to
the 4-gram model, was achieved when combining several language
models together by using the new bin estimation method. In the
speech recognition experiments, 4 \% reduction to the word error
and 7 \% reduction to the phonem error was achieved.
[Edit]
|
