VoxKnesset: A Large-Scale Longitudinal Hebrew Speech Dataset for Aging Speaker Modeling

2026-03-01

AI Summary

This work addresses the performance degradation of speaker verification systems due to age-related vocal changes, a challenge exacerbated by the absence of large-scale longitudinal speech datasets. To bridge this gap, the authors introduce VoxKnesset—the first large-scale Hebrew longitudinal speech corpus, comprising approximately 2,300 hours of parliamentary speeches from 393 speakers recorded over 15 years, accompanied by aligned transcripts and official demographic metadata. Benchmark evaluations using WavLM-Large, ECAPA-TDNN, and Wav2Vec2-XLSR-1B reveal that speaker verification equal error rates (EER) increase from 2.15% to 4.58% over the 15-year span. Furthermore, age regressors trained longitudinally successfully capture individual vocal aging patterns, whereas cross-sectional models fail to do so. VoxKnesset thus provides a critical resource for advancing research in longitudinal speaker verification and age modeling.

Technology Category

• Natural Language Processing: Speech
• Data Mining & Knowledge Management: Conversational Systems for Recommendation & Retrieval
• Machine Learning: Life-Long and Continual Learning

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• User Modeling, Personalization and Recommendation: Studies of user behavior, including longitudinal effects of personalized systems
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data

Abstract

Speech processing systems face a fundamental challenge: the human voice changes with age, yet few datasets support rigorous longitudinal evaluation. We introduce VoxKnesset, an open-access dataset of ~2,300 hours of Hebrew parliamentary speech spanning 2009-2025, comprising 393 speakers with recording spans of up to 15 years. Each segment includes aligned transcripts and verified demographic metadata from official parliamentary records. We benchmark modern speech embeddings (WavLM-Large, ECAPA-TDNN, Wav2Vec2-XLSR-1B) on age prediction and speaker verification under longitudinal conditions. Speaker verification EER rises from 2.15\% to 4.58\% over 15 years for the strongest model, and cross-sectionally trained age regressors fail to capture within-speaker aging, while longitudinally trained models recover a meaningful temporal signal. We publicly release the dataset and pipeline to support aging-robust speech systems and Hebrew speech processing.

Problem

• aging speaker modeling
• longitudinal speech dataset
• voice variation with age
• speaker verification
• speech processing

Innovation

• longitudinal speech dataset
• aging speaker modeling
• Hebrew speech processing
• speaker verification
• age prediction