Do you need to redact the audio, or only the transcript?

The short answer: transcripts by default#

The default for licensing call recordings is to deliver redacted transcripts and keep the audio in house. A transcript carries much of what buyers use from business calls: the problem described, the questions asked, the resolution and the outcome. Audio adds tone, pacing, accents and background sound, and it adds the speaker's voice, which is far harder to de-identify than text.

Audio belongs in scope only when the buyer's use depends on sound, such as speech recognition, turn-taking, voice agents or call-quality models. Ask what the buyer will do with recordings, and write that permitted use into the license so the scope cannot drift later.

Transcript only or audio too: a decision table#

The buyer's stated use decides the format. When the use is vague, start narrow, because a dataset can grow in a later delivery but released audio cannot be recalled.

Getting the transcript right comes first#

Transcript redaction comes first in either path, because muted audio is only as good as the transcript that located the spans. A transcript-only delivery also lives or dies on that pass, so it deserves the same care you would give a ticket or email export.

Call transcripts carry identifiers that ticket scanners rarely see. Agents introduce themselves by first name at the start of every call, speaker labels may hold agent names or extensions, and callers read out account numbers, addresses and order references in fragments. Replace speaker labels with consistent role tags such as Agent and Caller, and apply the same person placeholders used elsewhere in the dataset.

Speech-to-text errors cut both ways. A misheard surname may slip past detection, while a misheard product name can be redacted by mistake. Review a sample of the worst-quality calls, not only clean ones, since noisy lines and heavy crosstalk are where both kinds of error cluster.

If audio is in scope, how do you redact it?#

Audio redaction works from a time-aligned transcript. Speech-to-text output with word-level timestamps shows where each name, address, account number or card number was spoken, so those spans can be muted in the recording. Without timestamps, redaction becomes manual listening, which is slow and easy to get wrong.

Spelled-out names and numbers read in pieces are the classic misses. Transcription may render them as single letters or scattered digits that a PII detector does not recognize, so reviewers should listen for them specifically.

• 1Transcribe with word-level timestamps and speaker labels.
• 2Run PII detection on the transcript, then review a human sample.
• 3Map each finding to its time span, with a small pad on each side so partial syllables do not survive.
• 4Replace each span with silence or a tone, and use one choice across the whole dataset.
• 5Re-transcribe the redacted audio and confirm the removed words no longer appear.
• 6Listen to a sample, focusing on spelled names, numbers read digit by digit and crosstalk.

The voice is an identifier too#

The voice itself identifies a person, even when every name has been muted. Coworkers, customers and family members recognize voices, and voiceprint technology can match them at scale. Where recordings are used to create voiceprints or similar templates, biometric privacy laws may apply, so counsel should assess them before audio is scoped.

The Illinois Biometric Information Privacy Act is the best-known example. It lets a prevailing party recover liquidated damages of $1,000 per negligent violation or $5,000 per intentional or reckless violation, and in Rosenbach v. Six Flags (2019) the Illinois Supreme Court held that a person need not show actual injury beyond the violation to sue. A 2024 amendment limits recovery to one violation per person when the same identifier is collected repeatedly by the same method, but the exposure remains significant.

Each technical option changes the data. Pitch shifting is weak and often reversible. Voice conversion, which re-synthesizes speech in a different voice, protects identity better but alters the acoustic detail some buyers want. Excluding certain speakers, such as employees who were never told about this use, is sometimes the simplest route. Agree the approach with the buyer and record it in the license.

Consent, disclosures and the rest of the file#

Recording disclosures were usually written for quality assurance and staff training, not for licensing to an AI developer. Some states require consent from every party to a recording, and the wording of the original disclosure matters. Pull the IVR announcement text, agent call scripts, employee recording acknowledgments, the phone or transcription vendor's terms and any customer contracts that mention recordings, and have counsel compare them with the planned use.

Payment details need their own check. PCI DSS Requirement 3.3.1 says sensitive authentication data, which includes card verification codes, is not retained after authorization, so calls where a customer read out a security code are a payment-security problem before they are a licensing one. Exclude them rather than relying on muting.

Audio files also carry material a transcript pass does not catch. Give each item below its own check.

• 1Caller ID, phone numbers and customer names in file names, folder paths and call detail records.
• 2Voicemail greetings that state a person's name or direct number.
• 3Keypad tones entered during payment, which can reveal card digits.
• 4Hold music and recorded messages, which may be third-party content.
• 5Background voices, such as a television, a child or a coworker on another call.

Illustrative: an industrial distributor's inside sales calls#

Illustrative: a fictional industrial distributor records inside sales and customer service calls in its cloud phone system and stores transcripts alongside them. A developer building a parts-ordering assistant asked for the call history.

Counsel reviewed the recording disclosure and the buyer's stated use. The assistant needed conversation flow and product vocabulary, not voices, so the company licensed redacted transcripts with speaker turns and kept every recording in house. Customer names, account numbers and spoken card details were removed, and call detail records were reduced to date, duration band and outcome. A later request for audio was parked until the buyer could explain why sound was needed.

How SourceX scopes recordings#

SourceX scopes recordings by settling the transcript-or-audio question first, before any recording is prepared. Rights covers recording disclosures, employee notices and any biometric questions, and Preparation covers transcript redaction and, only where the supplier agrees, muted audio and voice treatment.

The permitted use in the SourceX Evidence Packet states whether audio is included and what it may be used for. The privacy record holds the padding rule, the silence or tone choice and the re-transcription results, so the buyer can confirm what was removed and the supplier can answer later questions without replaying recordings. The supplier approves both before Delivery.