For teams storing voice-call transcripts

Your caller said their SSN out loud.Now it’s sitting in a transcript.

On a call, people speak their card, phone, and social as words — ‘four zero seven, five five five…’ — which a document PII scanner reads as plain text and misses entirely. This redactor normalizes spoken digits, masks every span, and then re-scans its own output to prove nothing survived. When it can’t, it says so.

Not legal or compliance advice. This is a deterministic redaction aid with an honest self-check — not a certified de-identification product. It handles common spoken and written PII and tells you when it isn’t sure; specialized identifiers or free-form detail may need review. It scores no person, and a REDACTED CLEAN verdict is not a compliance guarantee. You remain responsible for what you store and share; confirm your obligations, including under HIPAA and state law, with counsel.

Get the Redactor — $89one-time · runs offline · uploads nothing
Five deliverables · runnable
Runnable Python redactor
stdlib
Verdict-reproducing workbook
xlsx
Findings log
csv
Redaction playbook
docx
Residual-review runbook
docx
The storage step before
Post-Call Intelligence Pipeline · Conversation QA Auditor
01.The Problem

A document scanner reads a spoken SSN as harmless words.

‘five two nine…’
a social read aloud — plain English words to a document PII scanner
re-scan
the tool checks its own masked output before it certifies anything clean
0 uploads
runs entirely offline; your transcripts never leave the machine
02.See It Work

Watch a spoken number get masked — and a split one get flagged.

Transcript in
Caller: Yeah it's four zero seven five five five zero one four two.
Agent: And your social?
Caller: It's five two nine, one one, four four three seven.
Agent: Thank you, Maria.
Redacted out
Caller: Yeah it's [PHONE].
Agent: And your social?
Caller: It's [SSN].
Agent: Thank you, [NAME].
1 phone · 1 ssn · 1 nameREDACTED CLEAN

The “spoken SSN + phone” call has no digits written anywhere — a document scanner sees only words and passes it clean. This tool normalizes the spoken digits first, masks them, then re-scans its own output. On the split-number call it admits a residual rather than certifying a transcript it isn’t sure about.

03.The Runnable Redactor

One command masks a folder of transcripts and logs every span.

The same masking you see in the demo comes from a zero-dependency Python redactor you run on your own transcript files. Here is its actual output on the shipped four-call sample:

SPOKEN-PII TRANSCRIPT REDACTOR
====================================================

PER-TRANSCRIPT
----------------------------------------------------
call-8801.txt            REDACTED CLEAN     email:1, name:2, phone:1, ssn:1
call-8802.txt            REDACTED CLEAN     address:1, card:1, dob:1, name:2
call-8803.txt            NOTHING TO REDACT  -
call-8804.txt            RESIDUAL EXPOSURE  name:2, phone:1
    RESIDUAL after masking: 1 span(s) still detected - review

BATCH ROLLUP
----------------------------------------------------
Transcripts .............. 4
Redacted clean ........... 2
Nothing to redact ........ 1
Residual exposure ........ 1
Total PII spans masked ... 13

VERDICT: REVIEW QUEUE

Masks text you provide and writes redacted copies. It uploads nothing
and scores no person. Not legal or compliance advice.
04.The Standard

Three commitments the redactor keeps on every transcript.

Spoken numbers are still numbers

A caller reads their SSN as ‘five two nine, one one…’. In the transcript that’s a string of English words a document scanner walks right past. This normalizes spoken digits back into a number, recognizes it, and masks it like any written one.

It checks its own work

After masking, it re-scans the redacted output. If anything is still detectable, it will not certify the transcript clean — it flags RESIDUAL EXPOSURE for a human. A redactor that might have missed something says so rather than pretending.

It claims only what it can guarantee

Names come from a list you supply, because blind name-detection produces both misses and false hits. The tool does the deterministic part exactly and hands the judgment part back to you — no confident guessing about your callers’ data.

05.What This Is — And Isn’t

A redaction aid with a self-check. Not a certified de-identifier.

It is
  • A deterministic redactor that masks PII and writes clean copies.
  • Spoken-digit aware — it catches numbers read aloud as words.
  • Self-checking — it re-scans its output and flags any residual.
  • Offline — it runs on your machine and uploads nothing.
It isn’t
  • A certified de-identification product — REDACTED CLEAN is not a guarantee.
  • A blind name-finder — you supply the known names.
  • A transcription tool — it works on transcripts you already have.
  • Legal advice, and it scores or ranks no person.

Not legal or compliance advice. This is a deterministic redaction aid with an honest self-check — not a certified de-identification product. It handles common spoken and written PII and tells you when it isn’t sure; specialized identifiers or free-form detail may need review. It scores no person, and a REDACTED CLEAN verdict is not a compliance guarantee. You remain responsible for what you store and share; confirm your obligations, including under HIPAA and state law, with counsel.

06.Who It’s For

Anyone whose call transcripts hold what callers said out loud.

Dental, med-spa, and clinic teams storing voice-agent transcripts
Agencies handling call data across regulated client accounts
Anyone feeding transcripts into a QA or analytics pipeline
Teams that need call records kept, but not the caller’s SSN with them
08.Common Questions

The parts worth pinning down before you buy.

Those flag what to redact — they explicitly do not remove or mask anything. This one performs the redaction: it masks every detected span and writes clean transcript copies. And it's built for voice, not documents: it catches PII spoken as words — 'four zero seven, five five five...' — which a document scanner reads as plain text and misses entirely. Different job (masking, not just flagging) and a different detection engine (spoken digits, not just written patterns).

Keep the call.
Not the caller’s social security number.

Mask what callers say out loud, prove it stuck, and store transcripts you can actually keep. One purchase, lifetime access, 12 months of updates. $89, once.

Not legal or compliance advice. This is a redaction aid with a self-check, not certified de-identification; a REDACTED CLEAN verdict is not a compliance guarantee. You remain responsible for what you store and share. Confirm your obligations — including under HIPAA and state law — with counsel.

Sold by RedHub AI LLC · Secured by Stripe · redhub.ai