Evaluating Audio Deepfake Detection Across Languages and Mobile Communication Channels

Abstract Audio deepfake detectors are frequently evaluated using curated speech corpora whose language, speaker, synthesis method and recording channel differ from operational audio. This narrative review critically examines how these differences affect the interpretation of detector performance, with particular attention to low-resource languages and mobile-mediated recordings. A purposive synthesis of 2024–2026 peer-reviewed journal literature links benchmark design, speaker and source separation, low-resource-language evidence, cross-dataset generalization, compression and acoustic stress testing. The evidence indicates that high discrimination on matched tests cannot establish reliability against unfamiliar generators, speakers, languages or transmission chains. We propose an evidence-informed evaluation workflow comprising documented data provenance; speaker-, source- and generator-disjoint partitions; matched baselines; isolated and combined distribution shifts; subgroup metrics with uncertainty; calibration and application-specific error analysis; and reproducibility reporting. Somali is discussed as an illustrative under-resourced case, not as a language for which the reviewed studies establish detection performance. Three original conceptual figures and three comparative tables translate the synthesis into a practical research and reporting framework. This framework is a methodological proposal, not a validated detector, standard or meta-analytic finding. Its use could make future evaluation of underrepresented languages more transparent and better aligned with real recording conditions.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-12
DOI
https://doi.org/10.21203/rs.3.rs-11250417/v1
Primary Topic
Speech Recognition and Synthesis
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Evaluating Audio Deepfake Detection Across Languages and Mobile Communication Channels

Abdikarim Abdi Abdulle, Abdullahi Salad Abdi, Omar Osman Haji Abdi, Maryan Mohamud Hambali
Research Square
Speech Recognition and Synthesis
preprint

Evaluating Audio Deepfake Detection Across Languages and Mobile Communication Channels

Abdikarim Abdi Abdulle, Abdullahi Salad Abdi, Omar Osman Haji Abdi, Maryan Mohamud Hambali
preprint en

Abstract

Abstract Audio deepfake detectors are frequently evaluated using curated speech corpora whose language, speaker, synthesis method and recording channel differ from operational audio. This narrative review critically examines how these differences affect the interpretation of detector performance, with particular attention to low-resource languages and mobile-mediated recordings. A purposive synthesis of 2024–2026 peer-reviewed journal literature links benchmark design, speaker and source separation, low-resource-language evidence, cross-dataset generalization, compression and acoustic stress testing. The evidence indicates that high discrimination on matched tests cannot establish reliability against unfamiliar generators, speakers, languages or transmission chains. We propose an evidence-informed evaluation workflow comprising documented data provenance; speaker-, source- and generator-disjoint partitions; matched baselines; isolated and combined distribution shifts; subgroup metrics with uncertainty; calibration and application-specific error analysis; and reproducibility reporting. Somali is discussed as an illustrative under-resourced case, not as a language for which the reviewed studies establish detection performance. Three original conceptual figures and three comparative tables translate the synthesis into a practical research and reporting framework. This framework is a methodological proposal, not a validated detector, standard or meta-analytic finding. Its use could make future evaluation of underrepresented languages more transparent and better aligned with real recording conditions.

Research Square
SIMAD University (SO)
Speech Recognition and Synthesis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon