Toward Interpretable Speech Deepfake Detection using Artifact-Specific Experts and Calibrated Detection Scores

Authors: Viola Negroni, Xin Wang, Wanying Ge, Paolo Bestagini, Junichi Yamagishi, Stefano Tubaro

Published: 2026-07-23 10:01:30+00:00

Comment: In peer review

AI Summary

This paper introduces an interpretable framework for speech deepfake detection that uses artifact-specific expert models. Each expert is trained to identify a particular speech synthesis artifact and produces an interpretable log-likelihood ratio as evidence. These individual scores are then aggregated to make a final real-versus-fake classification, allowing for transparent attribution of detection decisions to specific artifacts.

Abstract

In this work, we propose an interpretable framework for speech deepfake detection based on artifact-specific expert models. Rather than relying on black-box decisions, the framework provides human-understandable evidence, which is critical in high-stakes settings. Each expert is trained to detect a specific speech synthesis artifact, and its output is calibrated into a log-likelihood ratio that serves as an interpretable evidence score. We evaluate five artifact-specific experts and show that, with proper calibration, they can capture their target artifacts and produce meaningful evidence. Importantly, each expert estimates only the presence of its assigned artifact rather than directly performing the final decision. Their outputs are aggregated into an ensemble to produce the actual real-versus-fake classification, while maintaining interpretability by indicating how strongly each expert supports or contradicts a fake classification. Results show that artifact-specific experts capture interpretable signals of synthetic speech across multiple generation pipelines.

Key findings

The artifact-specific experts reliably detect their targeted artifacts, and proper calibration of their outputs is crucial for meaningful comparison and aggregation. Summation of expert LLRs proves to be the most effective aggregation strategy, providing an interpretable mechanism for identifying synthesis artifacts and diagnosing potential dataset shortcuts. The framework demonstrates generalizability to newer datasets by detecting persistent acoustic and phase distortions in modern architectures.

Approach

The proposed method trains multiple "expert" models, each specialized in detecting a specific speech synthesis artifact. These experts are trained using pseudo-fake data where the target artifact is intentionally introduced to real speech. The raw scores from each expert are then calibrated into log-likelihood ratios (LLRs), which are aggregated to produce a final real-versus-fake classification, providing human-understandable evidence for the decision.

Datasets

ASVspoof 2019, ASVspoof 5, SpoofCeleb

Model(s)

MLP (Multi-Layer Perceptron), Light CNN (LCNN)

Author countries

Italy, Japan