Some things we promise.

Your audio never leaves your account. We don't train models on it. We don't share it with other users. We don't sell it. You set how long it's kept — 7 days, 30 days, forever, or delete-on-end — in Settings → Memory.

Voice data for EU and UK users is stored in EU regions. Where a sub-processor (Deepgram for speech-to-text, OpenAI for the model) is US-based, we rely on Standard Contractual Clauses and the UK addendum. The current list with regions and data flows is at /legal/sub-processors.

Razen uses licensed, lifelike neural voices that run on your device. We do not clone real voices, and we never record, upload, or store a recording of your real voice. There is no "upload audio" field anywhere in the product.

In the free Voice Studio you can design your own synthetic voice by adjusting pitch, pace, and tone. Nothing you tune there is a recording of you — it's a synthetic voice, generated on-device, that isn't uploaded or stored on our servers.

• Watermarked. Every audio file we generate carries an inaudible C2PA signature plus container metadata identifying it as synthetic.
• No voice biometrics. Razen collects no voiceprints and no biometric voice data of any kind.
• Yours only. A voice you design in Voice Studio stays with your account and is never shared with other users.

Impersonating a real person is prohibited under our Acceptable Use Policy.

Some practice topics — grief, harm, addiction, family separation — carry real weight. We don't pretend otherwise. Scenarios marked sensitive show a warning before they start and a link to crisis resources in your region.

Razen is not a substitute for therapy, medical advice, or legal advice. If you are in distress, please reach out:

• UK — Samaritans, 116 123 — samaritans.org
• US — 988 Suicide & Crisis Lifeline
• EU — befrienders.org for country-specific lines

You can block whole categories of scenarios so they're never suggested to you again.

Razen is built for the EU AI Act's transparency obligations, in force from 2 August 2026.

• Article 50.1 (AI interaction notice). Every voice session begins with a short spoken statement that the partner is an AI and the conversation is synthetic. It cannot be skipped or customised.
• Article 50.2 (synthetic-content labelling). All generated audio is watermarked. Downloaded files carry an audible one-second announcement.
• Article 50.3 (no deepfakes of third parties). Razen does not clone real voices. Characters use licensed neural voices, and any voice you design in Voice Studio is fully synthetic — we never generate a deepfake of you or anyone else.

Razen is classified as a limited-risk system under the AI Act: we do not make automated decisions about people's access to employment, credit, education, or essential services. We are a coaching and practice tool.

We list every sub-processor publicly at /legal/sub-processors. The page is generated from a version-controlled file, so it's never stale. Today we use: Supabase (hosting, DB, auth, storage), Cloudflare (edge delivery), Stripe (billing), ElevenLabs (TTS), Deepgram (STT), OpenAI (LLM), Cartesia (TTS failover), Resend (transactional email), PostHog EU (analytics, consent-gated), Sentry (errors, no PII).

If you think someone is using Razen to impersonate a real person, write to razenai@outlook.com with as much detail as you can. We acknowledge within 72 hours. We hold the reported account within an hour pending review. Confirmed violations result in a permanent ban and, where appropriate, referral to law enforcement.

We also run a private bug bounty for security researchers. If you've found something, please write to security@razen.ai.