Request a demo

Research

Papers.

Working preprints on audio deepfake detection and the systems that run it. Each paper is readable in full on this site. None is peer reviewed, and none claims a venue, DOI or arXiv identifier, because none has one.

  1. 02Working preprint / 202614 pages

    Do generator artifacts survive a language change?

    A language-disjoint study of forensic representations for speech deepfakes

    Ayush Sahu, Tarini Sai Padmanabhuni, Sanjith Kumar

    Source tracing assumes a synthesis system leaves a trace that belongs to the system, not to the language it happens to be speaking. We test that by holding the generator fixed and replacing the language entirely: 62 generators learned across 12 languages, evaluated on 6 removed from the corpus before training.

    91.0%
    Top-1 on unseen languages
    3.2 pts
    Cost of the language change
    0.064
    Language entanglement ratio
    • Source tracing
    • Forensic embeddings
    • Cross-lingual control
    Read the paper
  2. 01Working preprint / 202625 pages

    Deploying an audio deepfake detector across GPU, CPU and Android

    Precision, runtime and platform in the deployment of an audio deepfake detector: a measured cross-target study

    Ayush Sahu, Tarini Sai Padmanabhuni, Sanjith Kumar

    Published work reports detection accuracy far more often than the cost of running the model. One 158.9M-parameter detector, seven inference configurations, three device classes, and the same 4,000-clip out-of-distribution set scored by every one of them.

    312x
    Sustained capacity range
    0.00040
    ROC-AUC range
    7 / 4,000
    Mobile vs server label changes
    • Deployment systems
    • Audio deepfake detection
    • Measured study
    Read the paper