SCAM DETECTED

⚠ Likely a scam call

    Report to the national cyber-crime helpline 1930 · cybercrime.gov.in
    VishGuard
    vishing & deepfake-voice detector · CC6-AI04 stt … llm … deepfake … ttl …

    Live microphone

    16 kHz mono, 3 s windows with 1 s overlap, streamed over WebSocket.

    idle
    The mic is off. Nothing is captured until you press Start.

    Analyse a recorded call

    wav · mp3 · m4a · ogg · webm — converted with ffmpeg, max 20 MB / 5 min.

    Paste a transcript instead

    Fastest way to test detection logic without spending speech-to-text credit.

    Attack simulator

    Scripted scammer dialogues written by the LLM and voiced with Sarvam Bulbul, then streamed through the real detector exactly as a live call would be.

    All audio is cached on disk, so a demo never depends on live API latency, rate limits or credits. Clips marked cached cost nothing to replay.

    Deepfake voice detection

    Two independent opinions: a wav2vec2 anti-spoofing model and a hand-built signal-feature scorecard.

    …

    Test your own clip

    Upload any voice recording to score it. Nothing is stored.

    Evaluation harness

    Labelled synthetic test calls run through the real pipeline. Every number on this page comes from a recorded run.

    These test calls are synthetic and LLM-written, then voiced by one TTS engine. Real scam calls have accents, crosstalk, codec loss and improvisation this set does not capture, so treat these numbers as an optimistic upper bound, not field accuracy.
    Download CSV Download JSON

    Add a real recorded call to the eval set

    The synthetic set is the weakest part of this project. Real audio makes the numbers mean more.

    Dashboard

    Live counters from this server process. Resets on restart — nothing is persisted.

    What is stored

    …

    Known limitations

    Populated from measured failures, not from a wish list.

    …