Internship on phonetic feature detection

Title: Automatic detection of phonetic distinctive features for speech recognition

Keywords: automatic speech recognition; pattern recognition; phonetics; audio indexing;

Automatic speech recognition heavily rely on the use of statistical models that hardly enable the integration of expert phonetic knowledge, in particular at the phonetic level. Recent work in our team have set up a theoretical framework to improve speech recognition systems based on the detection of broad phonetic landmarks (vowels, fricatives, etc.). The pre-detected landmarks are further used as constraints for a best path search algorithm.

The goal of the internship is to implement and evaluate technics for the automatic detection of broad phonetic features (broad phonetic classes, place and manner of articulation, voicing, etc.) using traditional pattern recognition methods (Bayes classifiers, SVM, neural networks, etc.). The work will first rely on an existing segmentation to evaluate various classification methods along with various signal features. The expected outcome of the internship is to determine which phonetic features can reliably be detected automatically, based on which method. However, the detection of broad phonetic features will also be evaluated whithin the entire automatic speech recognition system.

Skills: basic knowledge in pattern recognition and/or automatic processing of audio signal is required. Basic knowledge in speech processing or phonetics is welcome but not compulsory.

Supervision: Guillaume Gravier
Location: Irisa, équipe Metiss (http://www.irisa.fr/metiss)
Calification: Master
Duration: 4 - 5 mois
Salary: according to qualifications

Interested candidates should send a resume and a motivation letter to Guillaume Gravier.