Redact before anyone else reads
If you run QA, compliance or CX at a regulated business, you want to analyse calls, but every recording and transcript also holds card numbers, names and health details. The fix is to redact early, before analysis, sharing or any model use. Treat the audio and the transcript as two separate things to protect. Redacting the transcript replaces the text with placeholders, but the original audio stays unchanged, so the card number can still be heard by anyone with access to the recording. By the end, you can pick a redaction method for each data type and prove it worked on your own calls.
Match each data type to a method
Use this table as it stands, then adjust the rows to your own call types.
| Data type | Why it is risky | Best detection method | What to do to the audio |
|---|---|---|---|
| Names | Identifies a person on its own | Entity model | Silence the span |
| Phone numbers and emails | Direct contact route, easy to misuse | Pattern rules | Silence the span |
| Street addresses | Locates a person | Entity model, plus street-form patterns | Silence the span |
| Card numbers, expiry, security code | Direct fraud risk | Pattern rules with a Luhn check | Pause the recording, or never record |
| Bank account and ID numbers | Fraud and identity theft | Both: patterns plus context words such as "routing number" | Tone or silence |
| Dates of birth | Often used to verify identity | Both: a date pattern plus context such as "born on" | Silence the span |
| Health details | Sensitive, with no fixed format | Entity model, then human review | Silence the span |
| Passwords and security answers read aloud | Unlocks accounts | Context rules, since there is no fixed format | Never record |
Structured strings suit pattern rules. Card numbers, phone numbers and emails follow stable formats, so a pattern is precise, fast and easy to explain. Names, addresses and organisations have no fixed shape, so they need an entity model that reads context. Most production setups combine the two and reconcile the results.
Spoken digits need one extra step. A caller says "four two four two", and a pattern written for "4242" will not match it. Transcribe first, normalise the spoken numbers into digits, and only then run the rules. Pattern rules are fragile to spoken variation, and they miss numbers that were transcribed as words.
Detection happens on the transcript, and the matching audio spans are then replaced with silence or a beep tone. Silence is cleaner. A tone shows a reviewer that something was removed. Pick one and use it consistently.
The rule of thumb: the more predictable the format, the safer the pattern rule. The more the meaning depends on context, the more you need an entity model and a human review.
How the pipeline actually works
Redaction runs in five steps, and each depends on the one before it. First, transcribe the call so that every word carries a start and end time. Second, run detection on the text, using pattern rules with checksums for structured strings and an entity model for names and places. Third, map each detected span back to audio, from where its first word starts to where its last word ends. Fourth, replace the transcript text with a typed placeholder such as a card, phone or person tag. Fifth, overwrite that audio slice with silence or a tone, keeping the duration so timestamps still line up.
Setups usually add some padding around each span. Word timestamps are rarely perfect, so padding stops the edge of a digit leaking through. They also merge adjacent spans so no sliver of a digit is left between words.
Every step inherits the errors of the one before it. Accents, background noise and overlapping speakers raise transcription errors, and a number the transcript gets wrong is a number the detector never sees. Audio redaction can only be as good as the transcript timing.
Pause and resume avoids that dependency for card data. The recording stops while the customer reads the card, so the digits are never captured. Its weakness is the agent. If they pause late or forget, the card number and security code land in the file. PCI guidance advises against relying on manual pausing as a compliance strategy. It also does nothing for names, addresses or any other data type, and agents still hear the card details, so their desktops stay in scope.
If you want to talk through redaction on your own recordings, book a demo.
Tune for risk, not for count
Decide which error you can afford for each data type, because over-redaction and under-redaction fail in different ways. Redacting too little leaves personal data in files you believed were clean. Redacting too much removes the context your analysis needed. Pushing detection toward catching everything usually costs precision, and pushing toward preserving context usually costs recall.
So set the dial per data type. Be aggressive on card and account numbers, where a false positive costs almost nothing and a miss is serious. Be selective on names, where meaning depends on who is speaking about whom.
An education research nonprofit with interview transcripts already uploaded worried that automated pattern matching might be too aggressive or too loose on proper nouns, so it planned a review before deciding whether to redact.
The general rule is to redact the narrow category you are actually obliged to protect, not everything that looks sensitive. Do it before files reach any analysis model. Buyers keep asking where their recordings live and which models touch them, and redacting first gives you a short answer.
For files already uploaded, run a one-off sweep and review a sample of the output. Then set a rule so every new file is redacted on the way in.
Audit the output, not the settings
A redaction setting that is switched on proves nothing. Test the files it produces, in this order.
-
Build a seeded test set. Record calls where you know every sensitive value, using your agents' real accents and noise conditions. Read digits slowly, quickly, in groups and with corrections. Add harmless lookalike numbers, such as order references, to expose over-scrubbing.
-
Check the transcript and the audio separately. Audio redaction depends on transcript alignment, so the two can disagree. To test the audio, run transcription again on the redacted file and see whether any sensitive value comes back.
-
Count misses and false positives apart. A miss is leaked data. A false positive is over-scrubbing. Set a pass bar for each, and weight misses far more heavily, because undetected data is the larger compliance risk. Trace every miss to its source: the transcription, the entity model or the timestamp mapping.
-
Sample real calls on a schedule. Use a larger sample in the first weeks and after any change to transcription or detection. Weight it toward payment calls, overlapping speech and low-confidence detections, and add a random slice for baseline coverage.
-
Check every downstream copy. Redaction in one place does not clean the others. List exports, alerts, summaries, CRM notes and backups, and test each one.
-
Keep an audit log. Record what was detected, where, what action was taken, when, and by which rule or user. Include the category and confidence, but not the sensitive value itself.
For a quick leak test, search the redacted transcripts and exports for obvious patterns. A 16-digit string, an "@" sign or a run of digits after the words "card number" should return nothing. A hit means the pipeline failed on a case you now know about.
A clean transcript search says nothing about the audio file. Always listen to the recording as well.
Questions people ask
Does redacting a transcript also remove the data from the recording?
Usually not. Some tools redact only the transcript and leave the stored audio as it was, while others mute the sensitive segment in both. Check the setting. Then ask whether the original audio is kept or auto-deleted, and who can open it.
Is pausing the recording enough for PCI?
Only for card data, and only if the pause is enforced rather than left to the agent. Manual pausing is widely treated as unreliable because agents forget or pause late. Automated pausing is better, but the wider contact-centre environment can still be in scope.
Can I redact calls I already stored?
Yes, as a batch. List every copy first, including exports and backups, because a sweep of one location leaves the others untouched. Then review a sample of the output before you trust the run.
Related reading: 25 Security Questions to Ask AI Call Center Software Vendors Before You Sign, AI Voice Agent Business Case Template for Presenting to Leadership and Contact Center AI Risk Assessment Checklist for Customer-Facing AI Agents.
Check the audio file first
If you lead QA, compliance or CX at a regulated business, do one thing before any rollout. Take a single redacted call, put on headphones and play the audio from start to finish. A transcript can look spotless while a card number is still audible, and listening is the cheapest test you have.

