FLEX-SR (A FLEXible new approach to automatic Speech Recognition) | University of Oxford

FLEX-SR (A FLEXible new approach to automatic Speech Recognition)

Professor Aditi Lahiri (Faculty of Linguistics, Philology and Phonetics)

Professor Aditi LahiriProfessor Aditi Lahiri
A key principle of the Featurally Underspecified Lexicon (FUL) model, built within the framework of linguistics, is that “phonological features” (the articulatory and acoustic properties that make one sound crucially contrast with another) should have rigorous definitions and a finite set should adequately cover all the distinctive sounds across the languages of the world. This makes it possible for the FUL concept to be used as the basis of a novel Automated Speech Recognition (ASR) engine, and this idea was explored and developed under the ERC-funded WORDS project and the related Proof of Concept grant.

The FUL ASR system examines the acoustic properties of a speech sound and extracts distinctive features from it. These features are then matched against a lexicon to identify the words used. The result is an ASR system that requires little training, and employs theoretical insights from the study of how the human brain processes speech to emulate this process
on the computer.

Based on these principles, a mobile phone language learning application was produced enabling second language learners to improve their pronunciation. Words and sentences spoken into the App are analysed, and specific feedback is given like a personal tutor to improve and correct mistakes.