Non-English Response Detection Method for Automated Proficiency Scoring System

This paper presents a method for identifying non-English speech, with the aim of supporting an automated speech proficiency scoring system for non-native speakers. The method uses a popular technique from the language identification domain, a single phone recognizer followed by multiple languagedependent language models. This method determines the language of a speech sample based on the phonotactic differences among languages. The method is intended for use with nonnative English speakers. Therefore, the method must be able to distinguish nonEnglish responses from non-native speakers’ English responses. This makes the task more challenging, as the frequent pronunciation errors of non-native speakers may weaken the phonetic and phonotactic distinction between English responses and non-English responses. In order to address this issue, the speaking rate measure was used to complement the language identification based features in the model. The accuracy of the method was 98%, and there was 45% relative error reduction over a system based on the conventional language identification technique. The model using both feature sets furthermore demonstrated an improvement in accuracy for speakers at all English proficiency levels.