> ## Documentation Index > Fetch the complete documentation index at: https://docs.parable.so/llms.txt > Use this file to discover all available pages before exploring further. # Security & privacy > What Parable stores, who processes it, for how long, and the controls your workspace has over all of it Parable is an always-on workplace microphone, so we hold ourselves to a simple standard: you should be able to see exactly what we store, exactly who touches it, and turn every retention dial down as far as your team wants — including to zero for audio. ## What we store, and for how long | Data | Where it lives | Retention | | ------------------------------------------- | ------------------------------ | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Raw conversation audio | Cloudflare R2 (object storage) | **Configurable: 0–30 days.** Default 30. At 0, audio is deleted the moment transcription completes. 30 days is a hard ceiling enforced by a storage-level lifecycle rule. | | Transcripts, summaries, speaker attribution | Postgres (PlanetScale) | **Configurable: 30 days – forever.** Default forever — transcripts are the product. Measured from when the conversation happened. | | Voice-test recordings (enrollment audio) | Cloudflare R2 | **Configurable: keep, or delete after setup.** With retention off, each recording is deleted as soon as speaker setup completes. | | Voiceprints | Postgres (PlanetScale) | Kept while the voice profile exists. A voiceprint is a numeric fingerprint derived from the voice test — **it is not audio, and it cannot be played back or reversed into speech.** | Workspace admins set all three dials in **Settings → Privacy**. Policies are retroactive by design: shortening one deletes existing over-age data within the hour, continuously enforced by a sweep cycle — no grandfathering. Anyone with manage access can also delete an individual capture at any time, which permanently removes its transcript, summary, speaker data, and audio. ## Who processes your data Parable deliberately keeps the list of systems that ever touch your audio short. **No third-party LLM (OpenAI, Anthropic, Google, or anyone else) ever receives your audio or transcripts** — summaries and titles are generated by an open-source model running on infrastructure we control. | Processor | What it receives | Their retention | | ------------- | ---------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------- | | pyannoteAI | Conversation audio for diarization + transcription; voice tests for voiceprint creation | Uploaded media auto-deleted within 48 hours, job outputs within 24 hours; never used for model training ([their policy](https://docs.pyannote.ai/data-retention)) | | Modal | Audio and transcript text, processed inside Parable-controlled containers running open-source models | Ephemeral — compute only, nothing stored | | Cloudflare R2 | Encrypted audio objects (storage) | Per the retention table above; 30-day lifecycle ceiling on audio | | PlanetScale | The Postgres database (transcripts, metadata, voiceprints) | Per the retention table above | | Render | Hosts the Parable backend | Application logs only — no audio | | WorkOS | Sign-in identity (email, name) | Identity data only — no audio or transcripts | | Resend | Transactional email delivery (login codes, invites) | Email metadata only | All data is encrypted in transit (TLS) and at rest. Audio handed to pyannoteAI travels via short-lived signed URLs and is removed from our processing scratch space immediately after each job. ## Consent and identity * **Nobody is identified without opting in.** Speakers are named in transcripts only if they recorded a voice test. Anyone who hasn't enrolled appears as an anonymous "Speaker N" — always. * **Enrollment is explicit.** The voice test includes a consent step and is recorded in the product, by the person themselves. * **Forget-me works.** Deleting a voice profile removes the voiceprint and the voice-test recording, and detaches the person's name from every past transcript — historical speech reverts to an anonymous speaker. * **Recording is visible.** The Parable device shows a light whenever ambient capture is active, and capture can be switched off per device in Settings or on the device itself. One note for US deployments: some states (including California) require all-party consent for recording conversations. We recommend a short workspace policy acknowledgment for your team plus a visible-notice practice for guests — our [guest notice template](/guests) is ready to use. ## Deletion semantics, honestly When retention (or a manual delete) removes data, rows are purged from our operational database and objects from storage immediately. Database backups age out on their own cycle within about two days — so "deleted" means gone from live systems instantly and from backups within \~48 hours. Audio objects have no versioning; storage deletion is final. ## The voice-cloning question The raw material for cloning someone's voice is stored audio. Parable's answer is architectural, not just contractual: 1. Set conversation audio retention to **Delete after processing** — voice audio then exists only for the minutes it takes to transcribe. 2. Turn off voice-test retention — enrollment recordings are deleted once setup completes. 3. What remains is text plus voiceprints, and a voiceprint cannot be turned back into someone's voice. With both dials at zero, no durable recording of anyone's voice exists anywhere in Parable — ours or our processors'.