Data Protection Impact Assessment
Key finding: Because audio is processed in volatile memory and deleted immediately — and no personal data remains on our servers after processing — all identified risks are reduced to Low after mitigations are applied. The processing may proceed under the documented legal bases and safeguards.
Section 1
Processing Description
What we process and why
| Purpose | Data Processed | Legal Basis (GDPR) | KVKK Basis (Turkey) |
|---|---|---|---|
| Audio transcription | Audio file (RAM only, deleted after processing) | Art. 6(1)(b) — Contract performance | Explicit consent |
| Account & billing | Pseudonymous user ID, balance, usage metadata | Art. 6(1)(b) — Contract performance | Explicit consent |
| Authentication | OIDC account ID (hashed, original discarded) | Art. 6(1)(b) — Contract performance | Contract performance |
| In-app purchase | IAP receipt from App Store / Play Store | Art. 6(1)(b) — Contract performance | Contract performance |
| Age confirmation | Self-declared age confirmation (18+) | Art. 8 GDPR / KVKK Art. 6 | Legal obligation |
| Crash reporting | Anonymised error reports (opt-in, PII-redacted) | Art. 6(1)(a) — Consent | Explicit consent |
| Feedback & problem reports | Free text (PII-stripped server-side) + pseudonymous user hash; report type and app diagnostics | Art. 6(1)(a) — Consent | Explicit consent |
Data inventory
| Data | Server Retention |
|---|---|
| Audio file | RAM only — deleted after transcription |
| Transcript text | Until client acknowledgment (24-hour server failsafe if no ACK received) |
| Pseudonymous user ID | Until account deletion |
| Account balance + usage metadata | Until account deletion |
| Email address | Transit only — not stored |
| IP addresses | Transit only — not logged |
| Crash reports (opt-in) | SafeScribe’s crash reporting endpoint — not shared with third parties |
| Feedback & problem reports (opt-in) | PII-stripped free text + pseudonymous user hash only — retained up to 24 months, then auto-purged |
For complete data inventory details, see Privacy Policy § Data We Collect.
Per-job metadata includes: audio duration (seconds), file size (bytes), word count, charge amount (USD), and processing timestamps. It contains no audio content, no transcript text, and no user-identifiable information.
Backup retention. A single daily backup of account records is maintained for service continuity. Each backup overwrites the previous one. Data deleted via account deletion is removed from live systems immediately and from the backup within 24 hours — no copy persists beyond that window.
Processing flow
1. User records or selects audio on device
2. Audio preprocessed on-device (80 Hz high-pass filter, loudness normalization to -16 LUFS (speech-optimized, not broadcast-compliant) — peak limiting, 16 kHz resampling, FLAC encoding)
3. Encrypted upload to SafeScribe servers (TLS 1.3)
4. Server processes audio in RAM — self-hosted, a powerful model from the Whisper family via faster-whisper / CTranslate2, no third-party API calls
5. Transcript returned with SHA-256 integrity checksum
6. Client verifies checksum, acknowledges receipt
7. Server deletes transcript and audio from RAM immediately
8. Transcript stored locally on device in AES-256 encrypted storage
Section 2
Necessity and Proportionality
- ✓Audio upload is necessaryserver-side AI processing provides higher accuracy than current on-device alternatives at the quality level SafeScribe requires
- ✓Authentication is necessaryrequired for per-user billing and job isolation
- ✓Crash reporting is proportionatePII is redacted before transmission; opt-in only
- ✓Data is minimisedaudio processed in RAM only, never written to disk
- ✓Retention is minimisedtranscripts deleted immediately on acknowledgment; 24-hour server TTL failsafe if client never acknowledges
- ✓No secondary useaudio is never used for model training or analytics
Data subject rights
All GDPR and KVKK data subject rights (access, rectification, erasure, restriction, portability, objection, and consent withdrawal) are exercisable in-app or by contacting privacy@safescribe.dev. Data subject rights and their implementation are detailed in our Privacy Policy § What You Can Do.
Section 3
Risk Assessment
Identified risks and mitigations
| Risk | Inherent | Mitigation | Residual |
|---|---|---|---|
| Audio contains sensitive personal data (health, legal, financial) | High | RAM-only processing; immediate deletion; no persistent storage; no third-party access | Low |
| Unauthorised access to transcript in transit | Medium | TLS 1.3 enforced in production builds; SHA-256 integrity checksum | Low |
| Server-side breach exposing audio or transcripts | Medium | No persistent audio storage; authenticated API; per-user job isolation; TTL failsafe | Low |
| Unauthorised access to local encrypted storage | Low | AES-256 encrypted containers; key in iOS Keychain / Android Keystore | Low |
| PII leakage through crash reports | Low | Pattern-based scrubbing of emails, phones, IPs, and tokens before sending to SafeScribe’s own crash reporting endpoint | Low |
| Cross-border data transfer | Medium | User-to-controller uploads are not a GDPR Chapter V transfer (no EU exporter — EDPB Guidelines 05/2021 v2.0, Example 1); the GDPR binds us via Art. 3(2), with Art. 6 consent as the basis. The onward route through Cloudflare’s edge is a transfer and is covered by Cloudflare’s DPA/SCCs. KVKK Art. 9 covers cross-border transfer for users in Turkey. | Low |
| AI producing inaccurate transcript of sensitive content | Low | Transcription is informational only; user reviews all output; no automated decisions | Low |
All residual risks are Low. The primary risk driver — sensitive audio content — is addressed at the architectural level: audio is never written to disk, never retained beyond processing, and never shared with third parties.
Section 4
Consent and Transparency
Users are presented with a granular consent screen before first use, consisting of four independent cards:
Audio Processing (required)
Explains RAM-only processing, immediate deletion, and that audio is never written to disk or shared with third parties.
Stored Data (required)
Explains what persistent data is kept: pseudonymous account ID, credit balance, and anonymous usage metadata.
Age Confirmation (required)
Confirms the user is 18 or older. Required cards must be accepted to proceed.
Crash Reports (optional)
Allows opting into anonymous crash reporting. Defaults to off. Can be toggled at any time from Privacy Settings.
- Accept and Reject buttons have equal visual prominence (EDPB consent guidelines on dark patterns)
- Users can withdraw consent and delete their account at any time from Privacy Settings
- This DPIA is reviewed annually or upon significant processing changes
- Published at this URL for public transparency
Section 5
Decision
Based on the assessment above, all residual risks are Low. The processing may proceed under the identified legal bases with the documented safeguards in place. No prior consultation with a supervisory authority is required.
Review Log
Version History
| Date | Version | Changes |
|---|---|---|
| March 2026 | 1.0 | Initial pre-launch DPIA covering audio transcription pipeline, consent screen design, risk assessment, and legal basis documentation |