Introduction of semantic information in speech recognition

Orateur : Stéphane Level

Date : le 12 septembre 2019 à 10h30 – C005

Résumé :

Automatic Speech Recognition (ASR) is a growing industry. Indeed, there is an increasing demand from the industry for recognition systems or voice commands. The industrial use of this technology requires to have reliable and performing methodology. Current automatic speech recognition systems only use acoustic and linguistic information. These systems can encounter some problems in real conditions of use. Indeed, in a noisy environment the acoustic model is not very discriminatory and unreliable. The langage model allows to generate syntactically correct sentences but takes into account only the close words, the long-term context of words is not taken into account. To adress this issue, we propose a new approach to integrate the semantic information in ASR system by using the embedding methods, such as Word2Vec and BERT.