EchoGuard: Ambient Voice Defense for Alexa+ & Ring

An ambient defense agent for Alexa+ and Ring that detects synthetic voice cloning, flags high-urgency scams via AWS Bedrock, and automatically locks smart deadbolts in real time.

Comment

Inspiration

In 2026, artificial intelligence gave everyone the power of voice cloning with just 3 seconds of reference audio. While this enables magical experiences, it has unleashed a terrifying wave of virtual kidnapping, grandparent impersonation scams, and smart home exploitation. Vulnerable populations—especially elderly family members living alone—are losing their life savings to panic-inducing calls claiming a grandchild was arrested.

As Amazon transitions Alexa into Alexa+, giving it autonomous agentic power to control Ring smart deadbolts, order goods, and execute financial transactions, ambient voice security can no longer be an afterthought. We asked ourselves: What if Alexa+ wasn't just a passive speaker, but an active, intelligent guardian protecting the home from audio exploitation? That vision became EchoGuard+.

What It Does

EchoGuard+ operates as an autonomous background sentinel for Alexa+ and Echo Show devices:

  1. Acoustic Deepfake Forensics: Inspects incoming voice streams for neural vocoder phase truncations, unnatural harmonic uniformity, and absence of biological pitch jitter.
  2. Bedrock Semantic Coercion Detection: Evaluates dialogue transcripts through AWS Bedrock (Claude 3.5 Sonnet) to flag psychological pressure tactics, manufactured legal emergencies (IRS/police impersonation), and unauthorized door unlock attempts.
  3. Autonomous Smart Home Quarantine: If high-confidence fraud is detected, EchoGuard+ autonomously locks Ring front deadbolts, disables Echo 1-Click Voice Purchasing, and seals entry points before physical or financial harm occurs.
  4. Multi-Channel Emergency Dispatch: Immediately sends cryptographic incident reports with audio forensics logs and transcripts to designated family caregivers via SMS, Push, and visual Echo Show alerts.
  5. Caregiver PIN Override: Provides a seamless 4-digit PIN verification workflow so legitimate family members can release quarantines instantly.

How We Built It

  • Model Context Protocol (MCP) Server: Implemented in Python FastAPI conforming strictly to the Anthropic/Amazon MCP Specification (2025-11-25+ Streamable HTTP & JSON-RPC 2.0), exposing 4 deterministic tools: verify_voice_authenticity, detect_coercion_intent, quarantine_smart_home, and dispatch_guardian_alert.
  • AWS Bedrock AgentCore: Orchestrates multi-turn intent reasoning and urgency extraction without hallucination.
  • Audio Forensics Engine: Custom signal processing pipeline analyzing zero-crossing rates, spectral centroids, spectral flatness, and high-frequency attenuation ratios.
  • Simulated Alexa+ Visual Canvas: A responsive, dark glassmorphic Echo Show 15 web application built with Vanilla CSS and modern Web Speech API capabilities, delivering 60fps audio waveform visualizations and real-time WebSocket telemetry.

Challenges We Ran Into

  1. Streaming JSON-RPC over HTTP Keep-Alives: Long-running acoustic analysis combined with LLM inference occasionally pushed HTTP timeout limits. We resolved this by engineering a dual-transport architecture that combines immediate MCP acknowledgment tokens with an adjacent real-time WebSocket telemetry feed.
  2. Pydantic v2 to AWS Bedrock Schema Translation: Bedrock's schema parser rejects nested $defs references produced by standard Pydantic models. We wrote a custom schema normalizer to auto-flatten and dereference parameters into compliant Bedrock tool specs.
  3. Minimizing False Positives: Ensuring everyday conversational banter or emotional voice messages aren't accidentally flagged required fine-tuning our composite Risk Index to balance acoustic biomarkers with semantic urgency.

Accomplishments That We're Proud Of

  • Building a full-stack, standards-compliant MCP server with 4 functioning tools from scratch in under 24 hours.
  • Creating an ambient interface that preliminary judges can test in under 60 seconds with 1-click evaluation scenarios or their own live microphone.
  • Submitting a thorough, 4-page Amazon Developer Friction Log documenting actionable platform improvements for Amazon's developer tooling.

What We Learned

Building directly on the new Model Context Protocol (MCP) taught us how powerful standardized agent-tool communication is. Instead of proprietary brittle integrations, MCP allows an Alexa+ agent to interact with specialized security tools as first-class citizens.

What's Next for EchoGuard+

  • Integrating on-device C++ ONNX acoustic inference directly on Fire OS and Vega OS hardware.
  • Partnering with telecom providers and local emergency dispatchers for automated incident reporting.
  • Expanding Ring camera computer vision to cross-reference voice biometric authenticity with front-door facial verification.

Built With

Share this project:

Updates

Submission history