HCFD: A Benchmark for Audio Deepfake Detection in Healthcare

Abstract

In this study, we present Healthcare CodecFake Detection (HCFD), a new task for detecting codec-fakes under pathological speech conditions. We release the HCFD benchmark, a pathology-aware dataset containing paired real and neural-audio-codec-synthesized speech across multiple clinical conditions and codec families. Our evaluations show that state-of-the-art codec-fake detectors trained primarily on healthy speech perform poorly on HCFD, highlighting the need for healthcare-specific modeling.

We compare diverse pretrained audio encoders and show that PaSST is the strongest single-representation baseline for HCFD. Building on that, we propose PHOENIX-Mamba, a geometry-aware framework that uses long-context sequence modeling, multiple localized evidence vectors, and prototype-based reasoning in hyperbolic space to capture heterogeneous codec-fake modes in clinical speech.

Experiments across depression, Alzheimer's disease, and dysarthria in both English and Chinese show consistent gains over AASIST and strong pretrained-model baselines. With PaSST, PHOENIX-Mamba achieves the best reported accuracies: 97.04, 96.73, and 96.57 on the English tasks, and 94.41, 94.40, and 93.20 on the Chinese tasks.

Key Contributions

• 1
  We introduce HCFD, a benchmark task for codec-fake detection in healthcare speech under pathological speech variability.
• 2
  We build the HCFD benchmark, pairing bona fide pathological recordings with codec-resynthesized counterparts across multiple clinical conditions and two languages.
• 3
  We benchmark standard codec-fake detectors and pretrained encoders, showing major transfer failures from healthy-speech deepfake benchmarks.
• 4
  We propose PHOENIX-Mamba, which combines selective state-space modeling, multi-evidence pooling, and hyperbolic multi-prototype reasoning.

Method

PHOENIX-Mamba is designed for healthcare codec-fake detection where codec traces and pathological speech variability interact. The core idea is to avoid collapsing an utterance into a single pooled vector and instead retain multiple localized evidences, then organize fake evidence in hyperbolic space with self-discovered prototype modes.

Training Objective

The framework uses one negative prototype, multiple positive prototypes, evidence pooling with M = 4, hyperbolic embedding dimension h = 128, and optimized classification plus geometry-aware regularization.

Results

Table 1: Generalization of prior codec-deepfake detectors to healthcare speech

AASIST trained on standard CodecFake data transfers poorly to pathological healthcare speech. Training on in-domain healthcare data helps, and wav2vec 2.0 features help further, but a large gap remains.

The healthy-speech deepfake setting does not transfer cleanly to pathological speech. Alzheimer's is especially difficult.

Table 2: PHOENIX-Mamba with PaSST achieves the strongest reported scores

The paper reports PHOENIX-Mamba gains across encoders, with the strongest final numbers coming from the PaSST setup.

PHOENIX-Mamba contributes large gains beyond the strongest single-representation baseline, especially under clinical variability.

Ablation Study

The paper's ablations show that sequence modeling, multi-evidence pooling, and hyperbolic reasoning all matter. Removing any of them hurts performance substantially.

The largest degradation comes from collapsing the utterance to a single evidence vector. Hyperbolic multi-mode reasoning also gives a clear boost over Euclidean reasoning.

Datasets

HCFD covers three clinical conditions in two languages and preserves official split protocols while generating paired codec-synthesized speech.

Audio Samples

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

The current repository only includes the manifest file. Audio playback will work once the referenced waveform files are added.

Citation

If you use HCFD, cite the paper as follows:

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