Compliance
PII redaction and anonymisation in speech corpora
How names, numbers and identifiers are handled in audio and transcripts, and the trade-off between redaction and model utility.
- Audio
- Beeped or re-recorded
- Transcripts
- Tagged or masked
- Metadata
- Pseudonymised IDs
- Level
- Set by your spec
The short version
- Prompt design is the first defence: scripts avoid eliciting real identifiers wherever the use case allows
- Spontaneous and conversational capture will still surface names, places, card numbers and phone numbers, so transcripts are screened
- Redaction levels range from tagging identifiers in the transcript through masking them to removing the corresponding audio
- Speaker identity is replaced by an opaque speaker ID throughout the corpus
- Over-redaction damages utility: an ASR model that never hears a name will not transcribe one, so the level is a specification choice, not a default
Why it matters in practice
The right redaction level depends on what the model must do. A contact-centre ASR system has to recognise card numbers and names, so removing them from training data guarantees failure in production; a general-purpose assistant usually does not need them.
We set the level with you at specification time and apply it consistently, with the QA report confirming coverage.
How it is operated
| Level | Transcript | Audio | Use when |
|---|---|---|---|
| Tagged | Identifier wrapped in a tag, text retained | Untouched | Model must recognise identifier types |
| Masked | Replaced by a class token | Untouched | Text pipelines must not store identifiers |
| Redacted | Class token | Region replaced with tone | Corpus will be redistributed |
| Synthetic substitution | Fictional identifier of the same shape | Re-recorded segment | Realism required without real data |
What you receive
- The relevant policy or template as a document, not a claim on a web page
- Programme-specific artefacts delivered with the corpus manifest
- Named contact for audit questions during and after the programme
- Written confirmation at close-out
If your legal or procurement team has a questionnaire, send it with the specification. It is faster to answer it once, up front, than to unblock a signed programme later.
Frequently asked
Should we redact everything?
Usually not. Blanket redaction removes exactly the patterns a production model needs to recognise. Choose the level per identifier class against the deployment.
How is PII found in the audio?
Transcripts are screened by native reviewers with pattern-based flagging as a first pass, and every flag is confirmed by a human before any audio is altered.
Are speaker names ever in the delivery?
No. Speakers appear only as opaque IDs; the name mapping stays in the consent records, which are not part of the training delivery unless you require them.
Does redaction cost extra?
Tagging and masking are minor add-ons. Audio-level redaction and synthetic substitution are priced per audio hour because they require a second pass.
Related pages
Send your compliance questionnaire with the spec
We answer procurement, legal and security questionnaires alongside the technical scope, in the same working day where we can.