A video recording can sit untouched for months. Its transcript can be searched, copied into a ticket, exported to a review set, quoted in an email, or used to make a decision within minutes. Once speech becomes text, the organisation is managing a second data path with its own access, retention, correction and deletion consequences.

Govern the transcript as an explicitly inventoried data object, even when the recording and transcript originate from the same call. Give it an owner, purpose, access policy, legal-hold behaviour, correction workflow, redaction rule and deletion test. Do not assign a universal number of days to every transcript. Set the period from the processing purpose, applicable law, contractual duties, legal holds and the organisation’s record schedule.

Why the transcript needs its own control record

A transcript is not automatically a separate legal record from its recording. Operationally, however, transcript systems can create a searchable and exportable artifact that moves through different tools and reaches different people. A schedule that names only “the call recording” does not answer how the text is searched, exported, corrected, redacted or deleted.

That difference changes the control questions. A recording might be accessible only to a small investigations team, while its transcript is exposed through a wider search interface. A video export may require a deliberate download, while a text excerpt can be pasted into a case file or sent to an external recipient. A transcript can also preserve an ASR mistake as if it were a confident statement unless the workflow records what was corrected and why.

Start with an inventory entry for every transcript-producing workflow. The entry should identify:

  • the call, meeting or interaction that produced the transcript;
  • the business purpose and lawful basis for processing;
  • the people and systems that can search, view, export or edit it;
  • the source recording and any derivative formats;
  • the retention trigger and deletion event;
  • the legal-hold process;
  • the correction, redaction and audit requirements; and
  • the jurisdictions and contractual rules that apply.

That inventory is the foundation. The next decision is whether the text contains personal or sensitive information.

Classify transcripts by what was said

A transcript is personal data when its content identifies or relates to people. That classification follows from what the text contains, whatever system produced it.

The classification should follow the content and purpose. A transcript is not automatically special-category data, but spoken content can reveal special-category information. Article 9 of the 2016 GDPR identifies data revealing racial or ethnic origin, political opinions, religious or philosophical beliefs, trade-union membership, genetic data, biometric data used for unique identification, health data, and sex life or sexual orientation as special categories of personal data.

That means a filename such as support-call-1842.txt is not a sufficient classification. The inventory should consider what people said, why the conversation was recorded, who needs access and what decisions will be made from the text. A customer-support transcript, an employee investigation and a regulated financial conversation can have different purposes and different controls even when they use the same transcription service.

Content classification also determines redaction. Remove or mask information when the organisation does not need it for the stated purpose, when access should be limited to a smaller audience, or when disclosure would create an avoidable risk. Preserve the original under controlled access when it is needed for evidence, investigation or legal hold; do not silently rewrite it.

Set retention from purpose and obligation

Retention starts with the reason for keeping the transcript. A quality-review transcript, an employment-investigation record, a regulated customer interaction and a litigation hold are different records-management problems. The schedule should state the trigger that starts the clock, the event that ends it and the exception that suspends deletion.

For organisations subject to the EU GDPR, Article 5(1)(e) of the 2016 Regulation requires personal data to be kept in identifiable form for no longer than necessary for the purposes for which it is processed, subject to the stated exceptions. The rule applies to personal data rather than transcripts specifically. It applies to a transcript when the transcript’s content identifies or relates to people.

Translate that principle into a schedule rather than a slogan. Record the processing purpose, define how long that purpose remains active, and document the event that ends the need to retain the text. If a legal hold, statutory duty or contractual obligation prevents deletion, record the hold and restrict ordinary disposal until the hold is released.

India requires the same discipline through the Digital Personal Data Protection Act, 2023 and its Rules. Retention and deletion decisions should be mapped to those primary texts, the stated processing purpose, applicable sector obligations and any preservation duty. The Act and Rules should be checked in their current official versions before a schedule is approved; a policy copied from an EU or US workflow is not a substitute for that review.

For US federal civil discovery, Federal Rule of Civil Procedure 34 covers electronically stored information, including writings, sound recordings, images and other data compilations, and permits production in a reasonably usable form. That rule does not create a general corporate-retention period. It does mean that searchable transcript text can become a distinct electronically stored information object in preservation and production workflows.

When a dispute is reasonably anticipated, the legal-hold process must identify transcripts as well as recordings. Suspending deletion of the video while allowing the associated text to expire can create an incomplete preservation record. The hold notice should identify the systems, custodians, transcript exports and derivative files that fall within its scope. For the video side of the same record, Evidentiary Video Recording: Timestamps, Retention, and Export That Hold Up Later covers the timestamps and export trail a preserved recording needs.

Accuracy requires a correction trail

ASR output is a machine’s representation of speech. It can contain a wrong name, missing negation, substituted term or misleading punctuation. The organisation should measure its own audio, languages, terminology and review process instead of applying an accuracy number from another environment.

For EU personal data, Article 5(1)(d) of the 2016 GDPR requires personal data to be accurate and, where necessary, kept up to date. It also requires reasonable steps to erase or rectify inaccurate data without delay.

Accuracy does not mean silently overwriting an evidentiary transcript. Use a controlled correction model:

  1. Preserve the original transcript when its provenance or evidentiary value matters.
  2. Record the correction, the person or system that made it, the time and the reason.
  3. Attach a supplementary statement or reviewed version rather than destroying the original.
  4. Restrict processing of the disputed passage while the correction is assessed.
  5. Propagate the correction to approved search indexes, exports and downstream records.

The interface should make the status visible. A reviewer needs to know whether text is unreviewed machine output, human-verified text, corrected text, legally held text or redacted text. The record should also preserve the relationship between the transcript and its source recording.

Redact for the audience and the purpose

Redaction is a governance decision implemented through technical controls. Identify the information that the recipient needs, remove or mask information that recipient does not need, and preserve the original where a lawful investigation or hold requires it. Keep an audit record of who applied the redaction, what rule was used and which version was released.

Do not treat redaction as a one-time text replacement. Search indexes, exports, cached files, case-management copies and analyst notes can preserve the same sensitive passage. A deletion or redaction test should trace the transcript through each approved destination and verify that the released copy no longer exposes the restricted content.

Recording architecture affects what can be redacted and how confidently it can be traced. The companion technical discussion, Composite vs Per-Track Recording: The Choice That Decides Whether You Can Redact, examines why the underlying recording structure matters when a later review requires selective treatment.

Privacy-preserving deployment can reduce the number of parties that receive raw media or transcript content, but it does not remove the need for governance. On-Prem AI Video: Running Inference Without Sending Media to a Third Party explains the deployment boundary; the transcript still needs a purpose, access policy, retention schedule and deletion workflow inside that boundary.

Make deletion observable

A retention policy is incomplete until deletion can be demonstrated. Define which event starts deletion, which systems must receive the instruction, how legal holds pause it and what evidence confirms completion. Include primary transcript storage, search indexes, exports, backups and manually created case copies where the organisation controls them.

Deletion should not be confused with hiding a transcript from the main interface. A record that disappears from one search screen but remains in an export repository has not completed its lifecycle. The control should distinguish ordinary expiry, legal-hold suspension, approved archival, restricted processing and final deletion.

Test the process with a real transcript that contains personal information and a known redaction target. Confirm that the authorised user can find the record, that an unauthorised user cannot, that the correction trail remains intact, that a hold prevents disposal and that release of the hold permits the scheduled action. Record the result as evidence for the control owner.

The decision is straightforward: inventory transcripts separately, classify them by content and purpose, retain them only for a documented reason, preserve originals when evidence requires it, and make correction, redaction and deletion auditable. The limits are equally important. GDPR principles do not supply a universal transcript period, US Rule 34 is not a general retention mandate, and India’s requirements must be checked against the current Act, Rules and sector context.

Bottom Line

Once speech becomes searchable text, it creates a practical data path that recording-only schedules can miss. Give the transcript its own inventory entry, access controls, legal-hold behaviour, correction trail, redaction workflow and deletion test.

Keep the original where evidence requires it, release reviewed or redacted copies for narrower purposes, and set retention from the purpose and applicable obligation. A transcript lifecycle is complete only when the organisation can show who could use it, why it was kept, what changed and how it was removed.

What's Next

Transcript controls sit alongside the wider rules for recording, storage, access and export. The companion guide, Video Recording Compliance Across the EU, US and India: What Each Regime Demands, compares the primary compliance boundaries for video across those three jurisdictions.

Use it to test whether the recording schedule, consent process and export workflow around the transcript are governed consistently.

FAQs

How long should call transcripts be retained?

Retain a call transcript for the period required by its processing purpose, applicable law, contractual duties, legal holds and record schedule. There is no universal retention period that applies to every transcript, so the schedule should state its trigger, end event and deletion exception.

Are call transcripts personal data?

They are personal data when their content identifies or relates to people. A transcript can also contain special-category information when its spoken content reveals matters covered by Article 9 of the 2016 GDPR; classification follows the content and purpose.

Should a transcript be retained separately from a video recording?

It should be identified explicitly in the inventory even when it is retained with the recording. Search, export, access and deletion controls can treat text differently from video, and US Federal Rule of Civil Procedure 34 recognises electronically stored information in forms including writings, sound recordings, images and other data compilations.

Can an organisation correct an inaccurate transcript?

Yes. Preserve the original when its evidentiary value matters, add a reviewed or corrected version, record the change and restrict processing of disputed text while it is assessed. Article 5(1)(d) of the 2016 GDPR requires accurate personal data and reasonable steps to rectify inaccurate data without delay, but it does not require an organisation to erase the original evidentiary record.

Does GDPR prescribe a retention period for transcripts?

No. Article 5(1)(e) of the 2016 GDPR establishes a storage-limitation principle for personal data: identifiable data should not be kept longer than necessary for its processing purpose, subject to stated exceptions. The organisation must translate that principle into a purpose-specific schedule.

How should sensitive information be redacted from a transcript?

Define the recipient’s purpose, mask information that the recipient does not need, preserve the original under controlled access when evidence or a legal hold requires it, and record the redaction and released version. Check indexes, exports and downstream copies as well as the primary transcript.

It can. When a transcript falls within the information relevant to a dispute or investigation, the hold process should identify the transcript, its source recording, exports and related systems rather than naming only the video file.