AI & Voice Processing Notice

1. Overview

Zeitora's companion features use AI to understand your messages and respond in a supportive, reminiscence-focused way. When you use voice, your speech is converted to text, processed by the AI, and optionally read aloud using synthetic voice.

2. Voice conversation flow

When you record a voice message on the Companion screen:

• Your browser captures audio locally using the microphone (with your permission).
• Audio is sent over an encrypted connection to our NestJS API.
• OpenAI Whisper converts audio to text (speech-to-text).
• Our AI context engine retrieves relevant memories and recent conversation history.
• OpenAI GPT generates a text reply.
• ElevenLabs converts the reply to spoken audio (text-to-speech).
• The transcript, text reply, and audio are returned to your browser for display and playback.

3. Typed messages

Typed chat follows a similar path without speech-to-text or ElevenLabs: your message is sent to our API, contextualized with memories and session history, and answered by OpenAI's language model. Spoken playback is not generated for typed-only messages unless we add that feature later.

4. What is sent to AI providers

Depending on your interaction, the following may be transmitted to OpenAI or ElevenLabs:

• Your message text or voice-derived transcript.
• Relevant memory titles and descriptions retrieved for context (not your entire library unless relevant).
• Senior profile context needed for personalization (e.g., preferences you have saved).
• Recent conversation turns in the active session.
• For TTS: the AI-generated reply text (ElevenLabs does not receive your raw voice recording).

5. Models and providers

Large language model: OpenAI (default: gpt-4o-mini).

Speech-to-text: OpenAI Whisper (whisper-1).

Embeddings for memory search: OpenAI text-embedding-3-small.

Text-to-speech: ElevenLabs (configurable voice; default multilingual model).

We selected Western-headquartered providers with published API privacy terms. We do not route senior conversation data to China-based model providers.

6. Model training

We use provider APIs intended for business applications. Under standard OpenAI API terms, data submitted through the API is not used to train OpenAI's models unless you opt into separate programs.

You should still avoid sharing information you do not want processed by third-party AI systems.

7. Accuracy and human oversight

AI responses are generated automatically and may contain errors, inappropriate content, or outdated information.

The companion is not a substitute for professional medical, legal, or mental health advice.

Caregivers and families should supervise use where appropriate, especially for users with cognitive impairment.

8. Your choices

You may use typed chat instead of voice to avoid sending audio recordings.

You can decline microphone permission in your browser; voice features will not work without it.

You may request deletion of conversation history by contacting privacy@zeitora.app (subject to retention rules in our Privacy Policy).

9. Updates

We may change models or providers to improve quality or compliance. Material changes will be reflected in this notice and our Privacy Policy.