Detail publikace

Automatic Language Identification System

ČERNOCKÝ, J. MATĚJKA, P. BURGET, L. SCHWARZ, P.

Originální název

Automatic Language Identification System

Anglický název

Automatic Language Identification System

Jazyk

en

Originální abstrakt

This paper presents the language identification (LID) system developed in Speech@FIT. The system consists of two parts: Acoustic LID determines the language directly on the basis of features derived from the speech signal. We have improved existing approaches by adding discriminative training of acoustic models. In phonotactic LID, speech is first transcribed by phoneme recognizer into strings or graphs (lattices) of phonemes. On these, language models are trained to capture statistics of sequences of phonemes. We have pioneered the use of so called îanti-modelsî for this task. All experimental results are reported on standard NIST 2003 data; comparison with other published results is favorable to our system.

Anglický abstrakt

This paper presents the language identification (LID) system developed in Speech@FIT. The system consists of two parts: Acoustic LID determines the language directly on the basis of features derived from the speech signal. We have improved existing approaches by adding discriminative training of acoustic models. In phonotactic LID, speech is first transcribed by phoneme recognizer into strings or graphs (lattices) of phonemes. On these, language models are trained to capture statistics of sequences of phonemes. We have pioneered the use of so called îanti-modelsî for this task. All experimental results are reported on standard NIST 2003 data; comparison with other published results is favorable to our system.

Dokumenty

BibTex


@inproceedings{BUT22285,
  author="Jan {Černocký} and Pavel {Matějka} and Lukáš {Burget} and Petr {Schwarz}",
  title="Automatic Language Identification System",
  annote="This paper presents the language identification (LID) system developed in
Speech@FIT. The system consists of two parts: Acoustic LID determines the
language directly on the basis of features derived from the speech signal. We
have improved existing approaches by adding discriminative training of acoustic
models. In phonotactic LID, speech is first transcribed by phoneme recognizer
into strings or graphs (lattices) of phonemes. On these, language models are
trained to capture statistics of sequences of phonemes. We have pioneered the use
of so called îanti-modelsî for this task. All experimental results are reported
on standard NIST 2003 data; comparison with other published results is favorable
to our system.",
  address="Brno University of Defence",
  booktitle="Sborník příspěvků z odborného semináře "Nové technologie v radiokomunikacích"",
  chapter="22285",
  institution="Brno University of Defence",
  year="2006",
  month="january",
  pages="1--6",
  publisher="Brno University of Defence",
  type="conference paper"
}