Getting started

Guardrails

Safety controls for transcription: content moderation, personal data redaction, and profanity filtering.

Guardrails are safety controls you switch on when a transcript is created. They help transcripts meet compliance and safety requirements before the text ever reaches your product: personal details can be redacted, profanity masked, and harmful content screened during processing.

All three guardrails are boolean flags on POST /v2/transcripts, default false, and can be combined freely with each other and with the analysis flags described in Speech Understanding.

The three guardrail flags

Flag	What it does	Where you see the effect
`content_moderation`	Screens the recording for harmful or sensitive content during processing.	A `content_safety` report object on the completed transcript response, `null` unless the flag was requested. For a human-readable review on top of it, run an audit prompt (see Review flagged content).
`pii_redaction`	Detects personal details in the speech and redacts them from the transcription output.	The transcript content you receive back. Downstream consumers of `text` never see the redacted details.
`profanity_filter`	Masks profane words in the transcription output.	The transcript content you receive back, safe for customer-facing display.

Enable guardrails on a transcript

POST/v2/transcriptsCreate a transcript with guardrails enabled

bash

curl -X POST https://api.realtimevoicekit.com/v2/transcripts \
  -H "Authorization: Bearer rtvk_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{
    "audio_url": "https://example.com/support-call.mp3",
    "content_moderation": true,
    "pii_redaction": true,
    "profanity_filter": true
  }'

Guardrails work with both audio sources: a public audio_url as above, or an upload_id from the signed upload flow (POST /v2/uploads/sign, then POST /v2/uploads/{upload_id}/complete).

How results surface

The transcript response keeps the same shape whether or not guardrails are enabled: every field is always present. Redaction and filtering change the transcript content itself, while moderation fills in a dedicated report field.

Redaction and filtering are applied in place. With pii_redaction or profanity_filter enabled, the text and words you receive are already cleaned. There is no separate unredacted copy to fetch through the API, which keeps the sensitive version out of your logs, caches, and downstream systems.
Moderation returns a report. With content_moderation enabled, the completed transcript includes a content_safety report object describing what the screening flagged, with the flagged text, labels, timing, and confidence values. The field is null when the flag was off, and also when the screening found nothing.
Guardrails compose with analysis. A single request can redact personal data, mask profanity, and still enable speaker_labels or summarization.

If you passed a webhook_url when creating the transcript, the API POSTs a {"event", "transcript_id", "status"} notification to it when the transcript reaches completed or error, so review pipelines can react without polling.

Data controls and deletion

Guardrails cover what goes into a transcript; deletion covers how long it stays. When a retention window ends or a user asks for erasure, remove the transcript outright.

DELETE/v2/transcripts/{transcript_id}Permanently delete a transcript

Deletion is permanent and returns 204 No Content. Combined with pii_redaction, this gives you both ends of a compliance story: personal details never enter the transcript text, and the transcript itself can be removed on request.

Next steps

Review flagged content

Audit transcripts with targeted prompts and build structured moderation reports.

Speech Understanding

Summaries, sentiment, chapters, entities, and free-form questions over transcripts.