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

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 Contributions

• 1
  We formalize the novel task of Elderly CodecFake Detection (ECFD) and introduce the ECF dataset, covering English and Chinese elderly speech.
• 2
  We benchmark SOTA CF detectors trained on previous benchmark CF datasets and show substantial performance degradation on elderly speech, revealing poor cross-demographic robustness.
• 3
  We show that multimodal FMs (LanguageBind, ImageBind) are best suited for ECFD compared to SOTA speech-only FMs (Wav2vec2, WavLM, Whisper).
• 4
  We propose BONSAI, a novel framework that leverages Jensen-Shannon Divergence for fusion of FMs, achieving a new SOTA average EER of 1.66% with LanguageBind + ImageBind.

Method

BONSAI (Bridging FusiON via JenSen-ShAnnon DIvergence) fuses representations from two foundation models for ECFD. Rather than relying on simple concatenation, BONSAI explicitly aligns the distributional characteristics of heterogeneous FM representations using Jensen-Shannon Divergence (JSD), encouraging the two representations to capture complementary yet consistent information.

Figure: The BONSAI framework. Each foundation model (FM) branch is processed by a 1D-CNN and max-pooling, flattened, and aligned via JSD before concatenation and classification.

Alignment Loss

Training Objective

The classification loss L_CE and the JSD alignment loss L_JSD are jointly optimized, with lambda = 0.65 controlling the balance between the two. BONSAI has 3.8M-4.02M trainable parameters depending on the FM representation dimensionality.

Results

Table 1: Generalization of SOTA CF detectors to elderly speech (zero-shot)

Models trained on the prior CF benchmark (Lu et al.) generalize poorly to elderly speech. EER nearly doubles on the elderly subset of E2 (TIS) compared to the younger subset, indicating age-related degradation beyond simple distribution mismatch.

EER (%, lower is better). E1: SeniorTalk, E2: TIS Corpus.

Table 2: In-domain training and evaluation on ECFD

Multimodal FMs (ImageBind, LanguageBind) consistently outperform speech-only FMs (Wav2vec2, WavLM, Whisper) under both AASIST and CNN downstream classifiers, supporting the hypothesis that cross-modal pretraining benefits ECFD.

EER (%, lower is better). Avg is the average EER across E1 and E2 (Elderly).

Table 3: BONSAI fusion vs. concatenation

BONSAI consistently outperforms simple concatenation across all FM pairs, with the largest gains for multimodal and speech+multimodal combinations. The best pair, ImageBind + LanguageBind, achieves a new SOTA average EER of 1.66%.

EER (%, lower is better). Avg is the average EER across E1 and E2 (Elderly).

Dataset and Codecs

The ECF dataset pairs authentic elderly speech with codec-resynthesized counterparts produced by 14 neural audio codec variants spanning two languages.

Audio Samples

Sample rows below are loaded from audio/manifest.js. Each row pairs a real elderly recording with its codec-generated counterpart.

Citation

If you use the ECF dataset or BONSAI, please cite the paper as follows:

Update page numbers in the citation once the proceedings version is finalized.