Media (images, audio, video)

Redact PII from a support call recording

Transcribe a support call, then bleep card numbers, dates of birth and the customer's name in the audio with Redact Video — it handles audio-only files.

Support calls are recorded as audio files. Before they are shared for training, card numbers, dates of birth and the customer's name must be bleeped.

Redact Video handles audio-only input: with Sensitive speech set to Bleep, it transcribes the call, finds the sensitive words and bleeps them. There are no frames to analyse, so clear Find and choose a transcription connection as the Speech provider. It needs FFmpeg on the server (the Docker image includes it; Vercel does not).

1. Transcribe (optional, reusable)

If you also want the text, transcribe first with word timings and pass the result on, so the call is transcribed only once:

Transcribe the call
{
  "type": "media_transcribe",
  "label": "Transcribe call",
  "config": {
    "integration_id": "<integration-id>",
    "source": "{{form.call}}",
    "diarize": true,
    "word_timestamps": true,
    "formats": ["txt"],
    "output_var": "transcript"
  }
}

2. Bleep the audio

Add Redact Video. Set Video (or audio) to {{form.call}}, clear Find, set Sensitive speech to Bleep, choose Spoken information types (for example Credit / debit card number and Date of birth) and add the customer's name under Spoken words and phrases. Put {{transcript}} in Transcript to reuse step 1.

Bleep sensitive speech
{
  "type": "media_redact_video",
  "label": "Bleep PII",
  "config": {
    "transcribe_integration_id": "<integration-id>",
    "source": "{{form.call}}",
    "targets": [],
    "audio": "bleep",
    "audio_info_types": ["credit_card", "date_of_birth"],
    "audio_terms": "{{form.customer_name}}",
    "transcript": "{{transcript}}"
  }
}

The bleeped copy is {{redacted_video_file}}; {{redacted_video.audio_segments}} lists what was bleeped and when (text masked).

3. Check and share

Run Save & Test with a sample call, listen to the copy, and check audio_segments. Speech recognition can miss or mishear words, so review a sample of calls before relying on it. Then attach {{redacted_video_file}} to an email or store it.