Speaker: Ismaël Bada
Date and place: November 19, 2020 at 10:30, VISIO-CONFERENCE
Abstract:
We propose a method for lexicon adaptation in order to improve the automatic speech recognition (ASR) of non-native speakers.
ASR suffers from a significant drop in performance when it is used to recognize the speech of non-native speakers, since the phonemes of the foreign language are frequently poorly pronounced by these speakers. To take into account this problem of erroneous pronunciations, we integrate non-native pronunciations in the lexicon and subsequently we use this augmented lexicon for speech recognition of non-natives speakers. To realize our approach we need a small corpus of non-native speech and its transcription. To generate non-native pronunciations, we take into account graphemes in the analyzed pronunciations, with a view to automatically generating rules for creating new pronunciations, which will be added to the lexicon. We present an evaluation of our method on a corpus of non-native French speakers, pronouncing words in english.
ASR suffers from a significant drop in performance when it is used to recognize the speech of non-native speakers, since the phonemes of the foreign language are frequently poorly pronounced by these speakers. To take into account this problem of erroneous pronunciations, we integrate non-native pronunciations in the lexicon and subsequently we use this augmented lexicon for speech recognition of non-natives speakers. To realize our approach we need a small corpus of non-native speech and its transcription. To generate non-native pronunciations, we take into account graphemes in the analyzed pronunciations, with a view to automatically generating rules for creating new pronunciations, which will be added to the lexicon. We present an evaluation of our method on a corpus of non-native French speakers, pronouncing words in english.
