Speech Intelligibility in Noise: Perception and Prediction Models

Topic: In noisy environments, our ability to spatially separate the voice we are trying to understand from competing sound sources can significantly improve intelligibility. This spatial unmasking is based on binaural listening, meaning we use both ears rather than the monaural signal in only one ear. Numerous models exist for predicting speech intelligibility. These models are crucial tools for evaluating and developing hearing aids and cochlear implants that restore some audibility to deaf people. They are also necessary to improve building accessibility for people with hearing impairments. Monaural intelligibility models cannot predict spatial unmasking. To address this issue, we recently* proposed a preprocessor that processes binaural signals into "enhanced" monaural signals (in that the advantages of binaural listening are then taken into account) which can then be used by these models. The aim of this internship is to evaluate the performance of this preprocessor by designing an experiment (a listening test) to compare speech intelligibility under standard binaural conditions with that obtained for monaural sounds processed by the preprocessor. It will also involve varying the signal characteristics that influence intelligibility: the spatial configuration of the sources (simulated for headphone listening), the room reverberation level, and the type of competing sources (stationary noise, amplitude-modulated noise, voice). If the preprocessor correctly accounts for all aspects of binaural listening, then listeners will produce the same intelligibility scores with binaural sounds and those processed by the preprocessor. If not, its limitations will need to be identified. There are several possibilities for extending this topic in a PhD: improving the preprocessor and testing it on new data (from the literature or new experiments set up in the laboratory); comparing it to other intelligibility models; and predicting intelligibility for people wearing hearing aids and/or cochlear implants. These possibilities will be discussed based on the student's profile and preferences. * Lavandier et al. (2026) “A binaural front end for speech intelligibility models: application to the hearing-aid speech perception index (HASPI)” Trends Hear. (https://doi.org/10.1177/23312165261422005) Type of work: experimentation with audio signal analysis and processing (MatLab), implementation of controlled listening tests, statistical analysis of the data produced (R), publication. Potentially, measurement of room impulse responses (acoustic measurements, MatLab). Pursuing this research into a PhD is possible; this would be discussed with the student to identify potential funding sources. Contact : Mathieu Lavandier (mathieu.lavandier@entpe.fr) Université de Lyon, ENTPE, LTDS, Rue M. Audin, 69518 Vaulx-en-Velin Cedex https://mathieulavandier.wordpress.com/ Affiliation/structure d'accueil en S1: UMR CNRS 8248 - Equipe Audition (D. Pressnitzer) du DEC (validé les années précédentes auprès de Daniel Pressnitzer et Christian Lorenzi).