Bridging the Age Gap: Towards Detecting Neural Audio Codec Synthesized Elderly Speech Deepfake

Authors: Orchid Chetia Phukan, Girish, Mohd Mujtaba Akhtar, Chi-Chun Lee

Published: 2026-06-19 20:45:55+00:00

Comment: Accepted to INTERSPEECH 2026

AI Summary

This study introduces the Elderly CodecFake Detection (ECFD) task and the Elderly-CodecFake (ECF) dataset, demonstrating that existing deepfake detectors generalize poorly to elderly speech. The authors propose BONSAI, a novel framework that uses Jensen-Shannon Divergence to fuse multimodal foundation models (LanguageBind and ImageBind) for improved ECFD performance. BONSAI achieves a state-of-the-art average Equal Error Rate (EER) of 1.66%, highlighting the effectiveness of multimodal fusion for this challenging task.

Abstract

In this study, we introduce the Elderly CodecFake Detection (ECFD) task and release the Elderly-CodecFake (ECF) dataset in English and Chinese. We show that state-of-the-art CF detectors trained on previous benchmark CF datasets generalize poorly to elderly speech, revealing a critical vulnerability. We further hypothesize and demonstrate that multimodal foundation models (FMs) such as LanguageBind (LB) and ImageBind (IB) are more effective for ECFD due to their exposure to elderly content during cross-modal pretraining. Motivated by prior evidence that fusion of FMs enhances downstream performance, we explore fusion of FMs for ECFD. To this end, we propose BONSAI, a novel framework that employs Jensen-Shannon Divergence as the fusion mechanism. BONSAI with the fusion of LB and IB achieves an average EER (%) of 1.66 and outperforms individual FMs as well as competitive SOTA baselines, establishing a new benchmark for the ECFD task.

Key findings

State-of-the-art CodecFake detectors trained on previous benchmarks generalize poorly to elderly speech, showing a critical vulnerability. Multimodal foundation models (LanguageBind and ImageBind) are more effective for ECFD than speech-only FMs. The BONSAI framework, which fuses LB and IB using Jensen-Shannon Divergence, achieves the best performance with an average EER of 1.66%, setting a new benchmark for ECFD.

Approach

The proposed BONSAI framework fuses representations from multimodal foundation models (LanguageBind and ImageBind) for elderly speech deepfake detection. It employs Jensen-Shannon Divergence as a novel fusion mechanism to align heterogeneous FM representations in a unified space, which is then fed into a fully connected network for classification. This approach explicitly encourages distributional alignment between the merged features.

Datasets

Elderly-CodecFake (ECF) dataset (English and Chinese), SeniorTalk, TIS Corpora.

Model(s)

LanguageBind (LB), ImageBind (IB), Wav2vec2, WavLM, Whisper, AASIST, CNN.

Author countries

Taiwan, India

← Previous