Robust language recognition via adaptive language factor extraction

摘要

This paper presents a technique to adapt an acoustically basedlanguage classifier to the background conditions and speakeraccents. This adaptation improves language classification ona broad spectrum of TV broadcasts. The core of the systemconsists of an iVector-based setup in which language and channelvariabilities are modeled separately. The subsequent languageclassifier (the backend) operates on the language factors,i.e. those features in the extracted iVectors that explain the observedlanguage variability. The proposed technique adapts thelanguage variability model to the background conditions andto the speaker accents present in the audio. The effect of theadaptation is evaluated on a 28 hours corpus composed of documentaries and monolingual as well as multilingual broadcastnews shows. Consistent improvements in the automatic identificationof Flemish (Belgian Dutch), English and French are demonstrated for all broadcast types.