Articles écrits par altman

Date: Friday, June 19, from 1:00 p.m. Place: room S5 Speaker: Hugo Leguillier Title: On Low-Bit Quantization Errors in Speaker Verification: Diagnostic and Mitigation Abstract: Although low-bit quantization provides practical means to deploy speaker verification on resource-constrained devices, its effects on speaker verification performance remain poorly understood. In this paper, we study uniform K-means quantization-aware training of ResNet-36 and ResNet-200 through joint layer-wise and score-level analyses. Our layer-wise analysis highlights fragile components and shows that score degradation is not fully explained by weight distortion alone. We identify a clear knee point at 2 bits, with larger score drift and harmful decision flips concentrated near the FP32 threshold. Our score-level analysis reveals where and how score errors emerge under extreme quantization. Building on these findings, we propose a calibrated multi-precision cascade that resolves most trials at 2 bits and escalates only ambiguous cases, achieving performance close to FP32 while preserving the efficiency benefits of low-bit inference with substantially lower compute and memory costs.

Date: Friday, June 12, from 1:00 p.m. Lieu : Salle S5 Speaker: Hugo Daumain Title: From Self-Supervised Speech Models to Mixture-of-Experts for Robust Anti-Spoofing Abstract: Recent advances in speech generation have significantly improved the naturalness of synthetic speech, making spoofing detection increasingly challenging. A key limitation of current anti-spoofing systems is their limited robustness to unseen synthesis methods. In this work, we transform a self-supervised speech representation model into a Mixture-of-Experts (MoE) architecture to improve generalization. Feed-forward blocks in selected encoder layers are replaced by multiple expert networks controlled by a layer-wise gating mechanism, allowing experts to capture complementary acoustic patterns while preserving the representations learned during self-supervised pretraining. We further analyze the architectural choices affecting the performance of this MoE conversion and investigate the activation behavior of the experts. The proposed approach is evaluated on 14 spoofing datasets and reduces the macro EER from 5.46% to 4.81%, corresponding to 11.9% relative improvement over the baseline.

Date: Friday, June 12, from 1:00 p.m. Lieu: Salle S5 Speaker: Orane Dufour Title: A Large-Scale Per-Speaker Analysis of Re-identification Risk in Speech Anonymization Abstract: Speech anonymization is commonly evaluated using average-case metrics such as the equal error rate, which can hide large disparities in re-identification risks across individuals. In this paper, we conduct a large-scale per-speaker privacy analysis using a linkability-based metric under a worst-case scenario. Nearly 5,000 speakers are evaluated across multiple anonymization systems, attacker architectures, and conversation lengths. While linkability scores are highly polarized at the speaker level, the sets of easy to re-identify and hard to re-identify speakers vary substantially across configurations. We show that no single factor explains speaker vulnerability. Instead, the re-identification risk emerges from the interaction between the attacker, the anonymizer, and the amount of available speech. These results challenge the notion of intrinsic speaker-level privacy risks and emphasize the need for evaluation protocols that are explicitly conditioned on the attacker and anonymizer.

Salle 5 – 12h00 Titre: Modélisation Extensible de Langues et d’Accents Résumé: Les modèles d’identification de langue (LID) à l’état de l’art fonctionnent de manière fiable pour une centaine de langues. Cependant, derrière le concept de langue se cache de nombreuses variations émanant d’accents et de dialectes divers. Il est tout simplement impossible d’annoter les données en prenant compte de toutes ces variations. De plus, la parole accentuée engendre un comportement inattendu des modèles de LID et très peu de données annotées existent pour pallier le problème. Ce manque d’annotation empêche aussi l’augmentation de donnée via la synthèse de parole accentuée. Ce séminaire aborde ces problématiques et propose une ébauche de solution fondée sur une collecte de données à grande échelle à partir de diffusions radiophoniques. Cela permet d’associer aux données des annotations indirectes sous forme de géolocalisations. Le séminaire explore ensuite le lien entre la robustesse aux accents et la capacité à modéliser des séquences. Enfin, nous montrons comment ces modèles permettent d’améliorer les modèles de LID, en particulier sur la parole accentuée, et de faciliter l’extraction automatique de données accentuées pour entrainer des systèmes de synthèse vocale.Bio: Matthew Wiesner est un chercheur à Johns Hopkins University et chercheur  at Plus d'infos