University of Edinburgh

Published purpose

We will develop automatic systems to evaluate the proficiency of non-native learners of French, considering primarily the acoustic-phonetic aspects of speech, rather than the lexical content. We will focus on measures such as fluency (measured by speaking rate), quality of pronunciation, and intelligibility. Higher-level measures of proficiency such as vocabulary sophistication, grammar and topic development will be left to NRC researchers.To evaluate our chosen measures, we will build our own systems for automatic speech recognition (ASR) for French speech, building in the capability to work well on non-native accents. Using our own system will allow us to access to the internal properties of the system, necessary to determine information such as word-level timings, word-level confidence measures, and the deviation of the user from standard pronunciations at the phonetic level, as well as detecting when the user is speaking.Using the extracted measures as input features, we will train downstream classifiers – such as deep neural networks – to predict human judgments of spoken language proficiency on a collection of recordings from language learners. We will further explore whether these classifiers can be trained in a language-independent way, using a collection from learners of English, for example.