Skip to Main content Skip to Navigation

Transcription automatique de langues peu dotées

Abstract : With the development of technologies operating in a multilingual context, portability of speech technologies, and in particular, speech recognition, is a key challenge. State-of-the-Art speech recognizers are typically trained on very large amounts of data, both transcribed speech and texts. Recently there is a growing interest in developing speech technologies for languages for which only small amounts of data, and little or no linguistic expertise are accessible. For such languages, high Out-Of-Vocabulary (OOV) rates, poor language model estimation, are major issues. After studying the impact on recognition performance for the different types of training data : speech material used to train acoustic models ; texts corresponding to the transcriptions of the speech corpus ; and texts collected from newspapers and newswires available on the Web, automatic word decompounding to reduce OOV rates was investigated, with application to two case of studies : Amharic, the official language of Ethiopia and Turkish. A baseline algorithm was enhanced in order to address the problem of increased phonetic confusability arising from word decompounding, by incorporating phonetic properties and some constraints on recognition units derived from prior forced alignment experiments. OOV rates were reduced by 30% to 50% and relative word error rate reductions up to 5% were achieved. The algorithm is relatively language independent and requires minimal adaptation to be applied to other languages
Document type :
Complete list of metadatas

Cited literature [59 references]  Display  Hide  Download
Contributor : Magali Roserat-Brilhac <>
Submitted on : Tuesday, September 6, 2011 - 4:47:13 PM
Last modification on : Wednesday, January 6, 2021 - 10:26:02 AM


  • HAL Id : tel-00619657, version 1



Thomas Pellegrini. Transcription automatique de langues peu dotées. Informatique [cs]. Université Paris Sud - Paris XI, 2008. Français. ⟨tel-00619657⟩



Record views


Files downloads