This master internship is part of the REAVISE project: "Robust and Efficient Deep Learning based Audiovisual Speech Enhancement" (2023-2026) funded by the French National Research Agency (ANR). The general objective of REAVISE is to develop a unified audio-visual speech enhancement (AVSE) framework that leverages recent methodological breakthroughs in statistical signal processing, machine learning, and deep neural networks in order to design a robust and efficient AVSE framework.

The intern will be supervised by Mostafa Sadeghi (researcher, Inria) and Romain Serizel (associate professor, University of Lorraine), as members of the MULTISPEECH team, and will benefit from the research environment, expertise, and computational resources (GPU & CPU) of the team.

Audiovisual speech enhancement (AVSE) is defined as the task of improving the quality and intelligibility of a noisy speech signal by utilizing the complementary information provided by the visual modality, i.e., lip movements of the speaker [1]. Visual modality is especially important in high-noise situations, as it is less affected by acoustic noise. Because of that, AVSE could be exploited in several practical applications, including hearing assistive devices. Numerous works have already studied the integration of visual modality with audio modality to improve the performance of speech enhancement. While the majority of audiovisual speech enhancement algorithms rely on deep neural networks and supervised learning, they require very large audiovisual datasets with diverse noise instances to have good generalization performance.

A recently introduced AVSE approach is based on unsupervised learning [2,3], where during a training phase, the statistical distribution of clean speech is learned from a clean audiovisual dataset. This is done using a deep generative model, e.g. variational autoencoder (VAE) [4]. Then, at test (inference) time, the learned distribution is combined with a noise model to estimate the clean speech signal from the available noisy speech observations.

An important element of AVSE is audio-visual feature fusion, which should robustly and efficiently combine the two modalities. Current fusion mechanisms used for unsupervised AVSE are based on simple feature concatenation, which is not effective, as it treats the two feature streams on an equal basis. In fact, the audio modality usually contributes more than the visual modality, but in general, their contributions should be robustly balanced and weighted. In this project, we are going to develop efficient feature fusion modules based on attention models [5], which have proven very successful in different applications. The designed fusion module is supposed to robustly and efficiently incorporate the potentially different uncertainty (reliability) levels of the two modalities. We will then evaluate its effectiveness for AVSE.

References:

Background in statistical (speech) signal processing, computer vision, machine learning, and deep learning frameworks (Python, PyTorch) is preferred.

• Subsidized meals
• Partial reimbursement of public transport costs
• Leave: 7 weeks of annual leave + 10 extra days off due to RTT (statutory reduction in working hours) + possibility of exceptional leave (sick children, moving home, etc.)
• Possibility of teleworking (after 6 months of employment) and flexible organization of working hours
• Professional equipment available (videoconferencing, loan of computer equipment, etc.)
• Social, cultural and sports events and activities
• Access to vocational training
• Social security coverage

About 500 euros/month