@inproceedings{patil-etal-2026-frozen,
title = "What Frozen Speech Embeddings Reveal About Audio Deepfake Detection Difficulty",
author = "Patil, Yashas Harshad and
Rawat, Krish and
Shanbhag, Anup",
editor = {Chen, Nancy F. and
Chen, Yiming and
Gao, Xiaoxue and
Li, Haizhou and
Schuller, Bj{\"o}rn and
Sisman, Berrak and
Yue, Xianghu},
booktitle = "Proceedings of the Workshop on Secure, Safe, and Privacy-Preserving Speech and Audio {AI} ({T}rust{A}udio 2026)",
month = nov,
year = "2026",
address = "Hengqin, Zhuhai, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.trustaudio-1.1/",
pages = "1--10",
ISBN = "979-8-89176-452-1",
abstract = "Audio deepfake detectors perform unevenly across synthesis systems. Can frozen speech representations expose structure behind this variation? We construct a simple direction between genuine and synthetic WavLM embeddings and characterize each system{'}s position and spread along it. Across 61 MLAAD text-to-speech system{--}language groups, projected spread is strongly associated with mean detector difficulty across three runs ($\rho=0.60$). It is the only feature surviving correction over 83 recorded tests, and leave-one-system-out ridge prediction gives $R^2=0.277$. The genuine{--}synthetic directions also differ markedly across corpora, revealing domain-dependent representation geometry. As an auxiliary utterance score, the centroid projection reduces MLAAD detector equal error rate from 27.2{\%} to 16.3{\%}, improving on both component scores. Supervised probe comparisons further identify where adaptation gains originate. Together, these results establish a compact representation-space diagnostic of attack-specific detector behavior; the informative statistic and benefits of score adaptation depend on the corpus and detector."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="patil-etal-2026-frozen">
<titleInfo>
<title>What Frozen Speech Embeddings Reveal About Audio Deepfake Detection Difficulty</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yashas</namePart>
<namePart type="given">Harshad</namePart>
<namePart type="family">Patil</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Krish</namePart>
<namePart type="family">Rawat</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Anup</namePart>
<namePart type="family">Shanbhag</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the Workshop on Secure, Safe, and Privacy-Preserving Speech and Audio AI (TrustAudio 2026)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Nancy</namePart>
<namePart type="given">F</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yiming</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaoxue</namePart>
<namePart type="family">Gao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Haizhou</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Björn</namePart>
<namePart type="family">Schuller</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Berrak</namePart>
<namePart type="family">Sisman</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xianghu</namePart>
<namePart type="family">Yue</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, Zhuhai, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-452-1</identifier>
</relatedItem>
<abstract>Audio deepfake detectors perform unevenly across synthesis systems. Can frozen speech representations expose structure behind this variation? We construct a simple direction between genuine and synthetic WavLM embeddings and characterize each system’s position and spread along it. Across 61 MLAAD text-to-speech system–language groups, projected spread is strongly associated with mean detector difficulty across three runs (ρ=0.60). It is the only feature surviving correction over 83 recorded tests, and leave-one-system-out ridge prediction gives R²=0.277. The genuine–synthetic directions also differ markedly across corpora, revealing domain-dependent representation geometry. As an auxiliary utterance score, the centroid projection reduces MLAAD detector equal error rate from 27.2% to 16.3%, improving on both component scores. Supervised probe comparisons further identify where adaptation gains originate. Together, these results establish a compact representation-space diagnostic of attack-specific detector behavior; the informative statistic and benefits of score adaptation depend on the corpus and detector.</abstract>
<identifier type="citekey">patil-etal-2026-frozen</identifier>
<location>
<url>https://aclanthology.org/2026.trustaudio-1.1/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>1</start>
<end>10</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T What Frozen Speech Embeddings Reveal About Audio Deepfake Detection Difficulty
%A Patil, Yashas Harshad
%A Rawat, Krish
%A Shanbhag, Anup
%Y Chen, Nancy F.
%Y Chen, Yiming
%Y Gao, Xiaoxue
%Y Li, Haizhou
%Y Schuller, Björn
%Y Sisman, Berrak
%Y Yue, Xianghu
%S Proceedings of the Workshop on Secure, Safe, and Privacy-Preserving Speech and Audio AI (TrustAudio 2026)
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, Zhuhai, China
%@ 979-8-89176-452-1
%F patil-etal-2026-frozen
%X Audio deepfake detectors perform unevenly across synthesis systems. Can frozen speech representations expose structure behind this variation? We construct a simple direction between genuine and synthetic WavLM embeddings and characterize each system’s position and spread along it. Across 61 MLAAD text-to-speech system–language groups, projected spread is strongly associated with mean detector difficulty across three runs (ρ=0.60). It is the only feature surviving correction over 83 recorded tests, and leave-one-system-out ridge prediction gives R²=0.277. The genuine–synthetic directions also differ markedly across corpora, revealing domain-dependent representation geometry. As an auxiliary utterance score, the centroid projection reduces MLAAD detector equal error rate from 27.2% to 16.3%, improving on both component scores. Supervised probe comparisons further identify where adaptation gains originate. Together, these results establish a compact representation-space diagnostic of attack-specific detector behavior; the informative statistic and benefits of score adaptation depend on the corpus and detector.
%U https://aclanthology.org/2026.trustaudio-1.1/
%P 1-10
Markdown (Informal)
[What Frozen Speech Embeddings Reveal About Audio Deepfake Detection Difficulty](https://aclanthology.org/2026.trustaudio-1.1/) (Patil et al., TrustAudio 2026)
ACL