If you're building or maintaining a VoIP platform that records calls in the EU or UK, here's the uncomfortable truth: that generic "this call may be recorded for quality purposes" announcement probably doesn't make you GDPR compliant.
Voice recordings are personal data. AI transcripts derived from them are personal data. Vocal patterns are biometric data. GDPR treats all of it accordingly — and "staying on the line" is not consent.
Let's walk through what your call recording architecture actually needs.
The Four Legal Bases (Pick One Per Use Case)
GDPR Article 6 gives you six possible legal grounds. For enterprise telephony, four are relevant:
- Consent — Article 6(1)(a)
Must be freely given, specific, informed, and unambiguous via affirmative action.
❌ "By staying on the line, you consent..."
✅ "Press 1 to agree to recording, or press 2 to continue without recording"
Key engineering implication: callers can withdraw consent mid-call. Your system needs to halt recording immediately — not at the end of the session, not in a batch job. Immediately.
- Legitimate Interests — Article 6(1)(f)
The most common basis for QA and training recordings. But you need a formal Legitimate Interests Assessment (LIA) with three documented tests:
Test Question
Purpose What specific commercial benefit does recording serve?
Necessity Is recording strictly required, or could a less intrusive method work?
Balancing Do your interests override the caller's reasonable privacy expectations?
Under Article 21, callers can object to recording under legitimate interests. Your system must pause recording or route to an unrecorded line.
- Legal Obligation — Article 6(1)(c)
MiFID II / FCA mandate recording of trade-related communications. No opt-out. 5–7 year WORM retention. This isn't optional — it's regulatory.
- Performance of a Contract — Article 6(1)(b)
Verbal trade confirmations, verbal agreements. Recording is indispensable to completing the contract the caller requested.
The Retention Problem
GDPR Article 5(1)(e): keep recordings only as long as strictly necessary. Indefinite storage is unlawful. Period.
Different call types need different TTLs:
Category Retention
QA & training 30–90 days
Dispute resolution 1–6 years (statutory limitation)
MiFID II / FCA 5–7 years (WORM)
When TTL expires, deletion must be programmatic and permanent — not a soft-delete flag. Cryptographic shredding (destroy the encryption key, not the file) is one pattern.
Legal hold exception: compliance officers need to freeze specific recordings during audits/litigation, suspending auto-deletion without breaking the pipeline.
site: https://www.ecosmob.com/blog/gdpr-call-recording-rules/
The SIP Architecture Checklist
This is where it gets practical. Three areas your media pipeline must handle:
- SIPREC Media Forking with Metadata
Use SIPREC (RFC 7865) to fork audio from your SBC (Kamailio, RTPEngine, etc.) to centralized recording servers.
Critical: the fork must carry structured metadata:
Caller CLI (phone number)
Destination
Timestamp
Unique Call-ID
Without this, a Subject Access Request means grep-ing through terabytes of raw audio. That's not engineering. That's archaeology.
- DTMF Suppression (PCI-DSS + Data Minimization)
Configure media proxies to clamp/mute RFC 2833 DTMF tones before audio reaches:
Recording disks
Speech-to-text transcribers
This prevents credit card numbers and bank details from being stored in audio. Data you never captured is data you never have to protect, report, or delete.
- SAR-Ready Indexing
Every audio file and AI transcript must map to a unified customer entity identifier inside your Records of Processing Activities (ROPA, Article 30).
When a caller exercises:
Article 15 (right to access) → locate and export all recordings
Article 17 (right to erasure) → permanently delete across all stores
If your recordings aren't indexed to identities, you can't fulfill these rights. And "we can't find them" isn't a valid response to a supervisory authority.
The hard part isn't any single regulation. It's building one media pipeline that satisfies all of them simultaneously.
Breach Response: The 72-Hour Clock
GDPR Article 33: unauthorized access to stored recordings or transcripts → 72 hours to report to the supervisory authority (ICO, CNIL, etc.).
Voice recordings carry vocal biometrics → generally classified as high-risk breaches → mandatory public notification unless files were encrypted end-to-end at rest.
Before production:
Run a formal DPIA (Article 35) for automated recording, biometric matching, or AI sentiment analysis at scale
Appoint a DPO if processing large-scale systematic customer communications
Why Pilots Break in Production
The pattern I keep seeing:
Recording system works perfectly in demo
No unified customer identifier across audio files
No automated deletion pipeline
No DTMF suppression → credit card audio on disk
SAR request arrives → manual search through terabytes
Compliance becomes a fire drill
The cost of recording a call has never been lower. The cost of recording it wrong has never been higher.
TL;DR
Your GDPR-compliant call recording stack needs:
✅ Active opt-in (not passive disclaimers)
✅ Mid-call consent withdrawal capability
✅ Documented LIA if using legitimate interests
✅ Per-category retention TTLs with programmatic deletion
✅ SIPREC forking with structured identity metadata
✅ DTMF suppression before storage/transcription
✅ SAR-ready indexing across all recording stores
✅ End-to-end encryption at rest
✅ 72-hour breach reporting readiness
❌ "This call may be recorded" as your only compliance measure
Ecosmob builds compliant SIPREC pipelines, dynamic IVR consent workflows, and automated retention engines. Talk to their team about recording stack audits.
Top comments (0)