Audio and transcript redaction
How personal data is found and removed from call audio and transcripts.
Speech is transcribed locally inside the owner's environment. Personal data is detected in the transcript, located in time, and replaced in both places: a tone or silence in the audio, a tag in the text.
“Sure, you can reach me on [PHONE_NUMBER] any time after six.”
What is removed
| Entity | Tag |
|---|---|
| Names of people | [PERSON_NAME] |
| Phone numbers | [PHONE_NUMBER] |
| Addresses | [ADDRESS] |
| Card and account numbers | [CARD_NUMBER], [ACCOUNT_NUMBER] |
| Aadhaar numbers | [AADHAAR] |
| PAN | [PAN] |
| Policy and claim numbers | [POLICY_NUMBER], [CLAIM_NUMBER] |
| Email addresses | [EMAIL] |
| Dates of birth | [DOB] |
Numbers are caught whether they're spoken as digits, words, or a mix of languages, and when they're spelled out across several turns.
How it's replaced
- Audio. The span is replaced with a steady tone or with silence, chosen per dataset and stated in its datasheet. The rest of the call, including the speaker's voice, is untouched.
- Transcript. The span becomes a tag, and the manifest records each redaction's type, start and end.
How it's checked
Every batch goes through automated scans and a human review of a sample before the owner signs off. Results are in the de-identification report.
Voice
Redaction removes what people say about themselves, not their voices. Licences forbid any attempt to identify a speaker by voice or to link recordings to other data. See the permitted-use matrix.