Contributions à la Reconnaissance Automatique de la Parole Non Native

Abstract : Automatic speech recognition systems are still vulnerable to non native accents. Their precision drastically drops as non native speakers commit acoustic and pronunciation errors. We have proposed a new approach for non native ASR based on pronunciation modelling. This approach uses a non native speech corpus and two sets of acoustic models: the first set stands for the canoncial target language accent and the second stands for the non native accent. It is an automated approach that associates, to each phoneme from the first set of models, one or several non native pronunciations each expressed as a sequence of phonemes from the second set of models. These pronunciations are taken into account through adding new HMM paths to the models of each phoneme from the first set of models. We have developed a new approach for the automatic detection of the mother tong of non native speakers. This approach is based on the detection of discriminative phoneme sequences, and is used as a first step of the ASP process. As opposed to origin detection, we have proposed a multi-accent non native pronunciation modeling approach that takes into account several foreign accents. Besides, we have developed an approach of automatic phoneme-grapheme alignment in order to take into account the graphemic constraints within the non native pronunciation modeling. The gaol of this procedure is the sharpen the pronunciation modeling and enhance the ASR accuracy. We have studied some fast likelihood computation techinques, and we have proposed three novel appraoches that aim at enhancing likelihood computation speed without harming ASR precision.
Document type :
Theses
Complete list of metadatas

https://hal.univ-lorraine.fr/tel-01748528
Contributor : Thèses Ul <>
Submitted on : Thursday, March 29, 2018 - 11:37:29 AM
Last modification on : Tuesday, April 24, 2018 - 1:53:22 PM
Long-term archiving on : Friday, September 14, 2018 - 7:44:00 AM

File

SCD_T_2008_0101_BOUSELMI.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : tel-01748528, version 1

Collections

Citation

Ghazi Bouselmi. Contributions à la Reconnaissance Automatique de la Parole Non Native. Autre [cs.OH]. Université Henri Poincaré - Nancy 1, 2008. Français. ⟨NNT : 2008NAN10101⟩. ⟨tel-01748528⟩

Share

Metrics

Record views

54

Files downloads

13