Silent Echoes Compliance & Disclosures

What is captured

Agent microphone: Captured only after explicit in-product consent, only while coaching is active, and never while the call is paused or coaching is disabled.
Customer call-tab audio: Optional and separately consented. Coaching works on the agent's voice alone if it is declined or if capture fails.
Live QA transcription (opt-in): Only while an agent explicitly starts Live QA, short audio windows are sent to the AI transcription service, transcribed, and discarded. The resulting text is held in that browser tab for scoring and is never written to the database. Tone coaching itself never sends audio anywhere.
Nothing else: No screen contents, no keystrokes, no page text, no browsing history. The extension reads audio and nothing more.

Never stored: Raw audio and transcripts. Tone audio is analysed on-device; Live QA audio is transcribed in flight and discarded. Neither the audio nor the transcript is written to disk, to the database, or to any log.
Stored on your account: Session summaries (duration, average tone score, green share, peak streak, zone unlocks), state-change events with timestamps and cue text, your settings and calibration baseline, and cue thumbs-up/down ratings.
Locally only: Widget position, cached settings, and queued cue ratings live in extension storage on the agent's machine.

Per-user isolation: Row-level security scopes every session, event, setting, and rating to the account that created it. No agent can read another agent's data.
Admin role: Administrative rights live in a separate roles table that no client can write to; they are granted server-side only.
Agent controls: Pause a single call, disable coaching entirely, mute the mascot, reduce motion, and revoke consent at any time from Settings.
Kill switch: An admin rollout flag disables coaching across every installed extension, and a minimum version gate blocks outdated builds.

Beta accuracy: Tone scoring is a coaching signal, not a measurement. It should not be used for performance reviews, discipline, or compensation.
Voice diversity: Accents, soft speakers, and noisy floors shift scores. Calibration and the strict/normal/gentle sensitivity presets exist to compensate, and should be run per agent.
Analysis method: Scores come from on-device signal analysis of the voice — level, spectral warmth, turn-taking, expressiveness and steadiness. There is no speech recognition, no content analysis and no server-side audio processing. The web demo also offers a clearly labelled simulator for training and screenshots; sessions record which mode produced them.

Unlisted Web Store channel: Beta builds ship as a versioned package to an unlisted Chrome Web Store listing. Testers install once from a private link and Chrome auto-updates them; nobody re-loads unpacked folders, and every device converges on the same build.
Version gating: The admin console sets a minimum extension version. Builds below it stop coaching and show an update prompt, so a known-bad build can be retired within an hour without uninstalling.
Fleet kill switch and notices: Coaching can be paused for every tester at once from the admin console, along with a notice shown in the widget. Extensions poll hourly and on browser start.
Subject-access rights: Each agent can export a complete JSON copy of their data, or permanently delete all of it, from their own settings page — no administrator involvement required.

0 of 12 complete

Progress is saved in this browser so a reviewer can work through it in one sitting. It is a working aid, not legal advice.