AI Disclosure

VoiceNote AI uses two external AI providers to process your recordings: OpenAI for speech-to-text transcription and Anthropic for categorization and summarization. This page explains exactly what data is sent to each service and why.

All AI calls are made from our servers (Supabase Edge Functions), not from your device. Your OpenAI and Anthropic API keys are never included in the app — they live only in our server environment.

Service used: OpenAI Whisper API (whisper-1 model)

What is sent:

• Your raw audio file (M4A format), transmitted as a multipart upload directly to OpenAI's transcription endpoint.
• An optional language hint if you have set a preferred language in settings. If not set, Whisper auto-detects the language.

What is not sent: No account identifiers, email address, or metadata are included in the request to OpenAI. OpenAI receives only the audio content.

Why: Transcription is the core function of the app. Without sending the audio to a speech recognition model, no text can be produced.

OpenAI's data handling and retention policies for API inputs are described at openai.com/privacy.

Service used: OpenAI Embeddings API (text-embedding-3-small model)

What is sent:

• The AI-generated title and cleaned body text of your recording, concatenated into a single string (up to 8,192 tokens).

Why:Producing a numeric embedding of each recording's text allows the app to run semantic search — finding notes by meaning rather than exact keyword match. The resulting 1,536-dimension vector is stored in your account database and is not human-readable.

When you run a search, your search query text is also sent to the same API to produce a matching embedding, which is then compared against your stored embeddings entirely on our server. Your search query is not stored.

Service used: Anthropic Claude API (claude-haiku-4-5-20251001 model)

What is sent:

• The full verbatim transcript of your recording, exactly as returned by Whisper. No redaction or anonymization is applied.
• The current date in ISO 8601 format, so that Claude can interpret relative time references like "tomorrow" or "next Friday."
• Your device's UTC timezone offset, for the same reason.

What Claude returns: A structured JSON object containing category (task / note / journal), a short title, a cleaned body, optional due date, reminder preference, tags, sentiment, and mood score. Claude does not receive any prior recordings or account history.

The sentiment label and mood score are generated only for recordings classified as journal entries. These fields reflect the emotional tone of your own words and are stored solely for your personal use (displayed in the app and included in summaries). If you are in the EEA or UK, you may consider this data sensitive under GDPR. You can delete any recording — including its mood and sentiment data — at any time from the Library screen, or delete your account to remove all such data.

Why: Determining whether a recording is a task, a note, or a journal entry — and extracting dates, reminders, and labels from natural speech — requires a language model that can understand context and intent.

Anthropic's data handling policies for API inputs are described at anthropic.com/privacy.

Service used: Anthropic Claude API (claude-haiku-4-5-20251001 model)

When this runs: Only when you tap a daily, weekly, or monthly summary notification. If you have summary notifications disabled, this call never fires.

What is sent:

• The cleaned body text of all recordings from the summary period (day / week / month), concatenated into a single prompt.

What is not sent: Audio files, raw transcripts, account identifiers, or recordings from outside the summary period.

Why: Generating a coherent narrative summary of many separate voice notes requires a language model that can synthesize and prioritize across them.

• Your email address or account ID.
• The content of recordings other than the one being processed (except summaries, which explicitly send a batch).
• Your payment or subscription information.
• Your device identifiers or IP address (calls are proxied through our servers).

Transcription accuracy depends on audio clarity, accent, and background noise. Categorization and date extraction are probabilistic and may be incorrect — always review the output before acting on it, particularly for task due dates and reminders.

Questions about how AI is used in this app? Contact us at jamal.ib1810@gmail.com.