Speaking Faces for Face-Voice Speaker Identity Verification

Girija Chetty, Michael Wagner

    Research output: A Conference proceeding or a Chapter in BookConference contributionpeer-review

    4 Citations (Scopus)
    19 Downloads (Pure)


    In this paper, we describe an approach for an animated speaking
    face synthesis and its application in modeling impostor/replay
    attack scenarios for face-voice based speaker verification
    systems. The speaking face reported here learns the spatiotemporal
    relationship between speech acoustics and MPEG4
    compliant facial animation points. The influence of articulatory,
    perceptual, and prosodic acoustic features along with auditory
    context on prediction accuracy was examined. The results are
    indicative of vulnerability of audiovisual identity verification
    systems to impostor/replay attacks using synthetic faces. The
    level of vulnerability depends on several factors, such as the
    type of audiovisual features, the fusion techniques used for the
    audio and video features and their relative robustness. Also, the
    success of the synthetic impostor depends on the type of coarticulation
    models and acoustic features used for the
    audiovisual mapping of speaking face synthesis.
    Original languageEnglish
    Title of host publicationProceedings of the 9th International Conference on Spoken Language Processing Interspeech 2006 - ICSLP
    EditorsCarnegie Mellon
    Place of PublicationGermany
    PublisherInternational Speech Communication Association
    Number of pages4
    ISBN (Print)9781604234497
    Publication statusPublished - 2006
    Event9th International Conference on Spoken Language Processing - Pittsburgh, United States
    Duration: 17 Sept 200621 Sept 2006


    Conference9th International Conference on Spoken Language Processing
    Country/TerritoryUnited States


    Dive into the research topics of 'Speaking Faces for Face-Voice Speaker Identity Verification'. Together they form a unique fingerprint.

    Cite this