Communication Dans Un Congrès Année : 2011

Broadcast news speech-to-text translation experiments

Résumé

We present S2TT, an integrated speech-to-text translation system based on POCKETSPHINX and MOSES. It is compared to different baselines based on ANTS --- the broadcast news transcription system developed at LORIA's Speech group, MOSES and Google's translation tools. A small corpus of reference transcriptions of broadcast news from the evaluation campaign ESTER2 was translated by human experts for evaluation. The Word Error Rate (WER) of the recognition stage of both systems are evaluated, and BLEU is used to score the translations. Furthermore, the reference transcriptions are automatically translated using MOSES and GOOGLE in order to evaluate the impact of recognition errors on translation quality.

Fichier principal
Vignette du fichier
20110000-s2tt-us_letter.pdf (40.68 Ko) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence
Loading...

Dates et versions

hal-00628101 , version 1 (30-09-2011)

Licence

Identifiants

  • HAL Id : hal-00628101 , version 1

Citer

Sylvain Raybaud, David Langlois, Kamel Smaïli. Broadcast news speech-to-text translation experiments. The Thirteenth Machine Translation Summit, Sep 2011, Xiamen, China. pp.378-381. ⟨hal-00628101⟩
368 Consultations
333 Téléchargements

Partager

  • More