PhD defense Sicheng Mao: Generative Modeling for Robust Speaker Diarization

Monday 19 October, 2026, at 14:00 (Paris time) at Télécom Paris

Télécom Paris, 19 place Marguerite Perey F-91120 Palaiseau [getting there], amphi 2 and in videoconferencing

Jury

Laurent GIRIN, Professeur, Université Grenoble Alpes (Reviewer)
Lukas BURGET, Associate Professor, Brno University (Reviewer)
Jean-Francois BONASTRE, Professeur, Avignon University (Examiner)
Angélique DRÉMEAU, Enseignante-chercheuse, ENSTA (Examiner)
Roland BADEAU, Professeur, Télécom Paris (Thesis Supervisor)
Anthony LARCHER, Professeur, Université du Mans (Guest)
Mathieu FONTAINE, Maître de Conférences, Télécom Paris (Guest)

Abstract

Speech is a fundamental linguistic ability of human beings. In modern production and life, speech recognition is widely applied for information retrieval, speech understanding, and efficient human-computer interactions. As one of its crucial components, speaker diarization addresses the problem of « who spoke when » — it segments and clusters speaker-homogeneous segments based on their voice characteristics. This technology is widely applied in multi-speaker conversation scenarios, such as telephonic recordings, interviews, and meeting recordings, providing finer-grained annotation information for downstream speech-to-text transcription tasks.

In this thesis, we adopt generative modeling methods for speaker diarization for robustness against difficult acoustic environments, varying number of speakers, and overlapping speech.

Learn more

The state-of-the-art diarization model belongs to two main paradigms, namely the segmentation-clustering paradigm and the multi-label classification paradigm. For the former paradigm, we focus on the regularization of the deep speaker embeddings by a hierarchical variational autoencoder (HVAE) to improve its Gaussianity, which generalizes better to out-of-domain data and benefits the downstream variational Bayesian hidden Markov clustering model (VB-HMM). For the latter paradigm, we focus on the joint modeling of speaker diarization and source separation under multichannel settings. We propose a heavy-tailed separation model and a Bayesian diarization model with a Beta prior to improve the multichannel neural blind source separation and diarization model (Neural FCASA).

All the PhD defenses