August 19, 2026 · CensorMate Blog

How to Redact Audio: Removing Names & Personal Info from Recordings

Audio redaction is removing spoken information from a recording so it cannot be recovered — names, addresses, phone numbers, account details, medical information. It comes up anywhere recordings meet disclosure rules: public-records requests, court exhibits, HR investigations, research interviews under IRB consent, customer-call archives under GDPR. And it has a requirement casual censoring doesn’t: the redaction must be provable and irreversible.

Redaction is not bleeping

A censor bleep is editorial — it hides a word from listeners. Redaction is evidentiary — it must destroy the information. Three differences that matter:

1. Irreversibility. A bleep tone mixed over quiet dialogue can leave the original partially recoverable underneath. True redaction replaces the span with silence or tone at the sample level — the source audio in that window is gone from the output file, not covered.
2. Documentation. A records officer or paralegal has to answer “what was removed, where, and why.” Professional redaction produces a log: each redacted span, its timestamps, and the rule that matched it.
3. Chain of custody. For anything court-adjacent you also want file hashes of the input and output, so anyone can verify the redacted file is the one produced from that source.

The workflow (about 10 minutes per recording)

Step 1 — transcribe with word-level timestamps. Every modern redaction workflow starts with a transcript, because scrubbing a waveform by ear misses quiet mentions. This is also where the privacy problem starts: most transcription is a cloud service, and uploading an unredacted recording to a third party is itself a disclosure — often the exact thing your obligation forbids. On-device transcription sidesteps the whole question: CensorMate runs the speech-recognition model in your browser, so the recording never leaves the machine.

Step 2 — list what must go. Names and their variants (“Rob,” “Robert,” “Mr. Whitfield”), street addresses, employers, case-specific terms. Multi-word phrases work as single entries. CensorMate also auto-flags spoken patterns that look like phone numbers, email addresses, and SSN- or card-shaped numbers, and offers to redact them in one click — the numbers are the things human reviewers miss most.

Step 3 — review every match in the transcript. Redaction is a judgment call — the same surname might be a subject in one sentence and a public official in the next. A transcript with every match highlighted, where you toggle each one and can play each span in place, is the difference between redaction and hoping.

Step 4 — export with paperwork. In redact mode, CensorMate replaces each approved span with irreversible silence, then exports the redacted file plus a machine-readable log (CSV: each span, timestamps, matched rule) and a printable report with SHA-256 hashes of source and output — the chain-of-custody answer, generated automatically.

What about video?

Same pipeline — the audio track is rebuilt with redactions while the picture is untouched, which keeps body-cam footage, deposition video, and screen recordings intact as evidence. (Blurring faces/screens is a separate visual-redaction job; pair the two if you need both.)

Frequently asked questions

Is AI redaction defensible for legal or FOIA use?

The AI does the finding; a human approves every removal in the transcript review — the same division of labor as manual redaction, minus the missed mentions. What matters for defensibility is documentation (the log and hashes) and that the redaction is irreversible in the output file. Keep the unredacted original under your existing retention rules.

Why does on-device processing matter so much here?

Because the recording is, by definition, the sensitive artifact. Uploading it to a cloud redaction service creates a new disclosure to reason about (vendor agreements, retention, breach exposure). Processing in the browser means there is nothing to reason about: no upload path exists.

What does it cost?

Trying it is free — transcribe a recording (up to 10 minutes) and preview every redaction before paying. Exports: a $4.99 24-hour Day Pass (a one-time records request usually fits in a single pass) or $3.99/mo. Teams doing this regularly use Pro ($3.99/mo or $79 lifetime), which also removes the length limit for hour-long interviews and adds batch processing.

Can it redact filler or off-topic speech too?

Redaction targets your term list and detected PII patterns. For general cleanup (filler words, loudness) see how to censor podcasts — different job, same on-device pipeline.

Related: Audio & video redaction tool · Censor names & personal information · How accurate is the word detection?