Skip to content

PRIVACY_PLATFORM_GUIDE.md

Latest commit

 

History

History
97 lines (67 loc) · 3.18 KB

PRIVACY_PLATFORM_GUIDE.md

File metadata and controls

97 lines (67 loc) · 3.18 KB
Raw

Privacy, Platform, and Integration Guide

Privacy defaults

BargeKit is designed to be local-first.

  • Core state transitions run locally.
  • Synthetic fixtures are generated locally.
  • The demo is a static client-side page.
  • The browser microphone adapter only starts after an explicit user action.
  • This repository does not ship hidden upload, analytics, or recording flows.

What BargeKit does

  • Evaluates audio levels or external wake/push signals.
  • Detects speech segments using deterministic thresholds and timing windows.
  • Emits interruption, ducking, mute, noise-gate, and duplex-hold events.
  • Helps host apps keep UI state honest about whether input is really open.

What BargeKit does not do

  • Speech transcription
  • Text to speech
  • Wake-word model training
  • Cloud microphone streaming
  • Guaranteed full acoustic echo cancellation

Platform notes

Browser

  • Requires a user gesture before getUserMedia().
  • Browser/device echo cancellation quality varies widely.
  • AudioContext lifecycle and autoplay policies differ across Chrome, Edge, Safari, and embedded WebViews.

Laptop speakers + built-in mic

  • Expect more speaker bleed and false positives.
  • Start from the laptop_speakers preset.
  • Prefer stronger noise floor thresholds and slightly longer silence windows.

Wired / headset path

  • Usually allows a lower speech threshold.
  • Start from the wired_headset preset.
  • Half-duplex can often be disabled if the playback path is isolated enough.

Tuning recipes

Quick baseline

  1. Run synthetic fixtures.
  2. Confirm keyboard noise never crosses speechThreshold.
  3. Confirm long utterance opens speech reliably.
  4. Confirm interruption timing requests barge-in quickly while agent output is active.

Recommended starting presets

Profile speechThreshold noiseFloorThreshold debounceMs minSpeechMs silenceMs
laptop_speakers 0.64 0.22 90 140 520
wired_headset 0.50 0.14 70 110 420
quiet_room 0.44 0.10 60 100 380

VoicePath integration

Use BargeKit as the microphone/turn-taking layer and let VoicePath own output playback.

Suggested hooks:

  • bargekit.output.duck_requested -> lower output gain immediately
  • bargekit.output.cancel_requested -> stop or interrupt current speech turn
  • bargekit.user_speech.started -> stop queued agent speech if policy prefers user-first turns
  • bargekit.user_speech.ended -> decide whether to resume or regenerate output

createOutputDuckingController() helps avoid resuming stale output tokens after an interruption.

AgentPulse integration

createAgentPulseBridge() maps core events into a reducer-friendly stream:

  • user.speech.started
  • user.speech.ended
  • input.muted
  • input.noise_gated
  • barge_in.requested
  • input.half_duplex_hold

reduceAgentGlowState() turns those events into a minimal visual state model for listening / speaking / interrupted indicators.

Safe defaults for V1

  • keep processing local
  • require explicit microphone opt-in
  • expose mute and half-duplex truthfully in the UI
  • use interruption hooks, not magical echo-cancellation claims
  • document known limitations instead of hiding them