Audio Redaction Software

Nobody can scrub four hours of call-centre recordings looking for a card number. Audio has no thumbnail, no skim, no way to see where the sensitive part is.

So the practical route to a span is the transcript. Locating a phrase in text is faster and more precise than scrubbing the waveform, and the segment inherits its start and end from the transcript timing. That is the whole workflow, and it is why audio redaction at volume is a text problem wearing an audio costume.

1. Transcribe — speech is transcribed and aligned to the timeline, so every word is a jump point. Language mode is specific, auto-detect, or auto-detect multi-language.
2. Separate speakers — diarization distinguishes speakers and labels them sequentially, so the agent's side of a call can stay intact while the customer's is redacted.
3. Find — search the transcript for the words. Or let PII detection find them: names, contact details, government identifiers, financial identifiers, medical identifiers, plus country-specific formats for the US, UK, Spain, Italy, Poland, India, Australia and Singapore.
4. Replace — the segment is replaced with silence or a beep tone. Segments are also creatable by setting start and end times directly.

Spoken names, addresses, dates of birth, phone numbers, email addresses, government identifiers, card and bank details, medical identifiers, usernames, and any keyword or phrase you search for.

Transcripts are editable, so a mis-transcribed word can be corrected before it is used to drive a redaction.

This is where most vendors quote one number. There isn't one.

The figure that matters for redaction is the PII detection one. Transcription reaching 82 languages does not mean automatic PII detection does. In a language outside that set, the transcript still gets you keyword and pattern search, which is how multilingual recordings are handled in practice.

Bulk redaction runs across a set of files rather than one at a time, queued and worked unattended overnight. Bulk is permissioned per format, so an operator authorized for bulk audio is not thereby authorized for bulk video. Volume has been exercised at over 1.1 million recordings.

• Call recording redaction · 911 calls
• PCI DSS call recordings
• Call centers · Healthcare

• Segment boundaries follow transcript timing, so they are as precise as the transcript alignment.
• Redaction applies to the segments defined; accuracy follows the accuracy of those boundaries.
• Diarization distinguishes speakers within a recording. Attaching a name to a voice is speaker identification, a separate capability through a different provider.
• PII detection covers 10 languages. Transcription covers 82. Do not read the second number as the first.
• Transcription accuracy varies by language; per-language figures are published.

• Speech is redacted by marking segments and replacing them with silence or a beep.
• Segments are creatable by selecting words in the transcript, so locating what to redact does not require listening through the recording.
• Speakers are separated and labelled, so one party redacts while the other stays intact.
• Transcription covers 82 languages with published per-language error rates; PII detection covers 10.
• Transcripts are editable, under a separately licensed permission.
• Bulk audio redaction is a distinct permission from bulk video.