DESCRIPTION :
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), Romain Serizel (associate professor, University of Lorraine), as members of the MULTISPEECH team, and Xavier Alameda-Pineda (Inria Grenoble), member of the RobotLearn team. The intern will benefit from the research environment, expertise, and powerful computational resources (GPUs & CPUs) of the team.
Mission confiée
Generative models have increasingly become a fundamental tool in solving several inverse problems in an unsupervised way [1, 2]. This technique relies on the ability of generative models to learn the inherent characteristics of target clean data. Specifically, for speech enhancement, establishing a generative model acts as a data-driven speech prior, enabling the estimation of high-quality speech from noisy recordings without the direct need for corresponding pairs of clean and noisy data [2, 3, 4]. This unsupervised learning approach is particularly advantageous as it eliminates the dependency on extensive labeled datasets, which are often challenging and costly to procure, as done in supervised methods [5]. Moreover, training with only clean speech allows these models to better generalize to a variety of noisy environments they have never encountered, thus offering potentially broader applications in real-world scenarios where noise conditions are not predictable.
Principales activités
The use of variational autoencoders (VAEs) [3, 4] and diffusion models [2] represents the forefront of research in generative models for unsupervised speech enhancement. However, the field lacks a systematic comparison that evaluates these models side by side under standardized conditions. This project aims to bridge this gap through meticulously designed experiments that compare the effectiveness of VAEs and diffusion models in speech enhancement tasks. Each model will be implemented using similar network architectures to ensure that any differences in performance are attributed to the model capabilities and not to disparities in model complexity or configuration. The objective includes not only quantifying their performance in enhancing speech but also understanding their operational differences, resilience to various noise types, and computational efficiency. The insights gained from this analysis will provide valuable guidance for future developments in speech
processing technologies, aiming to optimize model selection and configuration for specific enhancement needs.
More precisely, the objectives of this project are outlined below:
* Implement both variational autoencoders and diffusion models using similar architectures to ensure comparability. Conduct detailed performance evaluations focusing on speech quality, intelligibility, noise reduction, and model efficiency under various noise conditions.
* Analyze the strengths and limitations of each model in handling diverse environmental noises and document their operational differences to determine their suitability for different speech enhancement scenarios.
References
[1] G. Daras, H. Chung, C.-H. Lai, Y. Mitsufuji, J. C. Ye, P. Milanfar, A. G. Dimakis, and M. Delbracio, A survey on diffusion models for inverse problems arXiv preprint arXiv :2410.00083, 2024.
[2] B. Nortier, M. Sadeghi, and R. Serizel, Unsupervised speech enhancement with diffusion-based generative models In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024.
[3] X. Bie, S. Leglaive, X. Alameda-Pineda, and L. Girin, Unsupervised speech enhancement using dynamical variational autoencoders IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 30, pp. 2993-3007, 2022.
[4] M. Sadeghi, and R. Serizel, Posterior sampling algorithms for unsupervised speech enhancement with recurrent variational autoencoder In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024.
[5] J. Richter, S. Welker, J.-M. Lemercier, B. Lay, and T. Gerkmann, Speech enhancement and dereverberation with diffusion-based generative models IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 2351-2364, 2023.
