dataspike-livekit-plugins 1.2.17

Dataspike deepfake detection plugin for LiveKit Agents

Dataspike Deepfake Detection Plugin for LiveKit Agents

This plugin integrates Dataspike with LiveKit Agents to provide real-time deepfake detection. It enables detection of synthetic or manipulated media in both video and audio streams during live or recorded sessions.

Installation

pip install dataspike-livekit-plugins

Prerequisites

You’ll need a Dataspike API key. Set it as an environment variable before running your agent:

export DATASPIKE_API_KEY="your_api_key_here"

Usage Examples

Basic Video Detection

from livekit.plugins import dataspike
from livekit.agents import AgentSession, Agent

async def entrypoint(ctx):
    await ctx.connect()
    session = AgentSession(...)
    detector = dataspike.DataspikeDetector()
    await detector.start(session, room=ctx.room)
    await session.start(agent=Agent(instructions="Talk to me!"), room=ctx.room)

Video + Audio Detection

from livekit.plugins import dataspike
from livekit.agents import AgentSession, Agent

async def entrypoint(ctx):
    await ctx.connect()
    session = AgentSession(...)

    # Configure video processing
    video_params = dataspike.VideoParams(
        burst_fps=1,            # Video FPS during suspicious state
        normal_fps=0.2,         # Video FPS during normal state
        quality=75,             # JPEG quality (0-100)
    )

    # Configure audio processing
    audio_params = dataspike.AudioParams(
        sample_rate=16000,      # Required: 16kHz PCM
        sample_size=48000,      # 3 seconds of audio
        interval=60,            # Minimum 60 seconds between samples
    )

    detector = dataspike.DataspikeDetector(
        video_params=video_params,
        audio_params=audio_params,
    )

    await detector.start(session, room=ctx.room)
    await session.start(agent=Agent(instructions="Talk to me!"), room=ctx.room)

With Voice Activity Detection (VAD)

from livekit.plugins import dataspike, silero
from livekit.agents import AgentSession, Agent

async def entrypoint(ctx):
    await ctx.connect()
    session = AgentSession(...)

    # Setup VAD for speech detection
    vad = silero.VAD.load()
    vad_stream = vad.stream()

    # Setup detector with audio processing
    detector = dataspike.DataspikeDetector(
        audio_params=dataspike.AudioParams()
    )

    # Start detector with VAD integration
    # Audio will only be processed when user is speaking
    await detector.start(session, room=ctx.room, vad_stream=vad_stream)
    await session.start(agent=Agent(instructions="Talk to me!"), room=ctx.room)

Configuration

VideoParams

Configure video processing behavior:

• burst_fps (float): Video sampling rate during suspicious state (default: 1)
• normal_fps (float): Video sampling rate during normal state (default: 0.2)
• quality (int): JPEG quality 0-100 (default: 75)

AudioParams

Configure audio processing behavior:

• sample_rate (int): Target sample rate in Hz (default: 16000, required by Dataspike API)
• sample_size (int): Number of samples per chunk (default: 48000 = 3 seconds at 16kHz)
• interval (int): Minimum seconds between audio samples (default: 60)

DataspikeDetector

Main detector configuration:

• api_key (str): Dataspike API key (or set DATASPIKE_API_KEY env var)
• video_params (VideoParams): Video processing configuration (default: VideoParams())
• audio_params (AudioParams): Enable audio detection (default: None, disabled)
• notification_cb (callable): Custom callback for detection events

Features

• ✅ Real-time video deepfake detection with adaptive frame rates
• ✅ Real-time audio deepfake detection with automatic resampling to 16kHz PCM
• ✅ VAD integration to process audio only during speech activity
• ✅ Automatic reconnection with exponential backoff
• ✅ Backpressure handling with bounded queues and frame dropping
• ✅ Customizable notifications via callbacks or room data messages

Links

• Dataspike API
• LiveKit Agents SDK