Prepare call recordings and transcripts for AI buyers.

Call data is valuable only if it can be cleared.

Contact centers, sales teams, and support lines hold years of recorded speech. Buyers of AI training data may want that material. Most of it cannot be sold as it sits today.

Recordings contain voices, names, account details, and sometimes health or payment information. Every call involves at least two people, and one of them is usually a customer. Your rights depend on what those people were told and what you agreed with them.

We represent sellers through that assessment and the commercial process that may follow. The first deliverable is a defined, documented subset. It is not a claim that the whole archive is licensable.

Consent and recording notices come first.

We start with the evidence of what callers and agents were told. Questions we build around your records:

• What recording notice played, and when did the wording change?
• Did it mention quality, training, analytics, or third-party use?
• Which countries or states were the callers and agents in?
• Do customer contracts restrict reuse, disclosure, or transfer?
• Do agent employment or contractor terms address recorded voice?
• Were any calls for regulated sectors, such as health or finance?
• Are retention or deletion duties still running on any segment?

Your counsel reviews the answers and the proposed scope. We keep a register of findings, exclusions, and open questions. Calls without adequate support are excluded, not argued into the package.

PII redaction and what remains.

Redaction is a processing step with its own record. We document the method, the entity types covered, and the residual risk.

A workable plan addresses:

• Names, phone numbers, addresses, and account identifiers in text
• The same details spoken in audio, which need muting or replacement at the timestamp
• Payment card details, which should be handled at the source where possible
• Voice itself, which can identify a person even when the text is clean
• Metadata fields such as agent IDs, call IDs, and timestamps

Redaction changes the data. Buyers will ask how it affects model quality, so we describe it plainly. Some buyers require redaction before sample access. Others will contract for controlled evaluation instead.

Speaker separation and alignment.

Buyers test structure before they test volume. Two checks matter most.

Speaker separation. Is each speaker on a separate channel, or mixed into one? Are labels consistent, such as agent and customer? Mono recordings need diarization, and we report its error rate on a reviewed sample.

Audio and transcript alignment. Do timestamps match the audio at word or segment level? Was the transcript produced by people or by a model, and which model? Are corrections tracked?

We also record format details: codec, sample rate, channel count, call duration, and silence or hold segments. Missing information is labeled rather than reconstructed as certainty.

Hebrew, Arabic, dialects, and code-switching.

Language coverage is more than a language field. For Hebrew and Arabic collections, we document:

• Modern Standard Arabic versus regional dialects, by call share where known
• Hebrew spoken registers and loanwords
• Code-switching between Hebrew, Arabic, English, and other languages within one call
• Transcription conventions, spelling variants, and script choices
• Domain terms specific to your sector

Estimates are kept apart from reviewed samples. If a dialect has not been checked by a qualified reviewer, we say so. Buyers building speech systems for these languages often value that distinction.

Sample hours and what buyers test.

A sample should represent the proposed scope, not the best calls. We select across queues, dates, agents, and languages, and we describe known gaps.

Buyers commonly test:

• Audio quality, noise, and clipping
• Transcript accuracy against the audio
• Speaker labels and timestamp alignment
• Redaction completeness
• Language and dialect distribution against the description
• Duplicates, empty calls, and unusable files
• Provenance and the documented rights basis

The sample package includes a data dictionary, a manifest, approved recordings, quality limitations, and evaluation conditions. Each version has an identifier. If counsel removes a category, counts and descriptions change before buyers see them.

Timeline and commercial terms.

Work begins once you can identify the source systems and provide the notice and contract documents. Rights review, redaction, and internal approvals are separate dependencies. There is no fixed closing period, and we agree stage dates after reviewing those inputs.

Buyer evaluation starts only after you approve materials, recipients, and access conditions. A license can limit permitted uses, reserve your continued use, or separate a historical export from later refreshes. Exclusivity should name the exact content and field of use.

Related reading: acquired archives, company shutdown data, and our insights. For the sequence of work, see how it works.

Start with a confidential brief.

Tell us what systems hold the recordings, roughly how much exists, and what callers were told. Use the confidential seller brief. We will review it and respond on next steps. Your counsel and the authorized owner decide whether to proceed.