← Back to blog
Call Center Datasets: What They Contain and How to License Yours

Call center agent handling a customer conversation

A call center dataset is a collection of call records, transcripts or audio files used to train, test or benchmark machine learning models. Public sets exist, but the biggest English one is restricted to non-commercial use. Developers who build commercial products therefore need licensed data, and companies that run support operations hold exactly that kind of data. This guide covers what these datasets contain, what the public ones allow, and what you should check before licensing your own.

NordicmartechPut Your Call Data to WorkMode connects companies holding valuable support records with AI developers, while you retain ownership and share only de-identified copies.Explore the data marketplace

TL;DR

Table of Contents

What a call center dataset is

The phrase covers three different things, and buyers and sellers often mix them up.

Call records (metadata). Call center phone systems log call identification numbers, actions and elapsed times. Computer telephony integration can add the reason for the call, agent actions and related account information, as an academic overview of call center data describes. This is tabular data: useful for staffing and performance analysis, thin for training a language model.

Transcripts. Text of the conversation, usually with speaker labels and sometimes topic or sentiment tags. This is what people mean by a call transcript dataset.

Audio. The recordings themselves, paired with transcripts. Audio is what you need to train speech recognition or separate speakers.

Industry guidance sorts datasets into transcript-only and audio-paired categories. The first suits language understanding, fine-tuning large language models and sentiment analysis. The second suits acoustic model training, speech recognition and speaker diarization.

Call center dataset types and their uses

What the public datasets offer

The public landscape is small, and most of what ranks for this keyword is either a teaching spreadsheet or a research release.

CallCenterEN

The most cited real-world example is CallCenterEN. According to its research paper, it holds 91,706 English call center conversations, with 10,448 hours of original audio duration. The transcripts are PII-redacted and delivered as JSON with metadata. The calls come from outsourcing partners, and the paper lists inbound and outbound calls.

The audio is not in the public release, because of biometric privacy risks. That detail matters. Even the researchers who built the set chose not to publish voices.

The authors say it fills a gap in publicly available real-world speech recognition corpora and supports customer-support and sales AI research.

CallCenterEN

The license problem

CallCenterEN uses a CC BY-NC 4.0 license. It is meant for academic research and non-commercial model development. Commercial use, resale and redistribution are prohibited.

If you are building a product that you will sell, you cannot train it on that set. This is the gap commercial buyers face, and it is why licensed, rights-cleared business data has value.

Teaching spreadsheets

Many results are small tabular files built for analytics practice. They hold fields like agent name, topic, resolution status and satisfaction score. They are good for learning dashboards. They contain no conversation text, so they do not help train a model to understand or generate support dialogue.

What makes a call center dataset valuable to AI developers

AI teams do not pay for volume alone. They look at a handful of features.

Feature Why a buyer cares
Real conversations, not scripted Models trained on staged calls miss interruptions, corrections and unclear requests
Speaker labels Separate agent from customer so a model learns each role
Topic or intent labels Allows supervised training and evaluation
Outcome fields (resolved, escalated, satisfaction) Lets teams train and test models against real results
Redaction quality Names, account numbers and card data must be removed reliably
Clear rights A buyer needs a license that permits commercial training
Consistent format JSON or similar structured delivery saves cleanup time

The last two are where most datasets fail. A transcript archive can be rich and still unusable if no one can show who consented to what.

Why your own records may be worth more than a public set

A public set is a snapshot from a few outsourcing vendors. A company that has run support for ten years holds something different: a sequence of customer questions, agent answers, ticket follow-ups and CRM notes about one specific industry. A finance firm has conversations about disputes and account servicing. A software company has support tickets, chat and code review history. A consultancy has client emails and documents.

That depth is hard to get any other way. It is also why AI developers look for business records that carry real-world context rather than another generic corpus.

Our guide to AI training data explains where such data comes from and how licensing works in more detail.

Checks to complete before you license call data

Support data carries more risk than most business records, because it holds the voices and personal details of your customers. Work through these checks with your general counsel before any data leaves your systems.

1. Consent and notice

The FTC has said that collecting and using consumers’ voice data without adequate consent would itself violate the FTC Act. Read your call recording notices and privacy policy. Do they describe use for AI training, or only quality and training of staff? If the notice is narrow, talk to counsel about whether the recordings can be used at all, or whether transcripts that are properly de-identified are the safer starting point.

State law adds layers. Recording consent rules differ between states, so counsel should review the states where your callers are located.

2. Payment card data

If agents take payments by phone, the payment card industry standard matters. The PCI Security Standards Council’s guidance on telephone-based payments says call centers may not retain card verification codes after authorization, including in digital audio recordings. The same guidance calls for disposal procedures and limits on how long card information stays on recording servers and CRM databases.

This guidance is from the version 2.0 era of the standard. Ask your compliance team to confirm the current requirements, then scan your archive for card numbers before anyone considers sharing it.

3. De-identification

Redaction has to cover more than names. Think account numbers, addresses, email addresses, phone numbers, order details and anything a caller says that could point to a person. Redaction works best at the transcription step, before downstream systems see the text. Test it on a sample and have a person read the output.

Voice itself is biometric, which is why the public CallCenterEN release left audio out. Treat text-only transcripts as the lower-risk starting point and decide on audio separately.

4. Leakage paths

The risk is not only the deal itself. Data also leaks through staging copies, shared drives and internal AI tools. A review of enterprise AI data leakage paths is a useful way to map where copies of your records sit before you prepare any of them for sharing.

5. Contract terms

Before signing, make sure the agreement covers:

Counsel should also check existing customer contracts. Some of them restrict use of conversation content.

Preparing the data

Preparation is mostly about structure. A buyer wants consistent files, so decide on the scope first.

  1. Inventory. List where calls, chats, tickets and CRM notes live and how far back they go.
  2. Choose a format. Transcripts as structured JSON with speaker labels, timestamps and metadata are easier to evaluate than loose text files. If your records start as web pages or documents, a pipeline for extracting transcripts into structured data shows how stages like fetching, extraction and normalization work in practice.
  3. Redact and sample. Run de-identification, then review a random sample by hand.
  4. Document provenance. Record when and how the data was collected and under which notice.
  5. Hold the originals. Only de-identified copies should ever be shared.

This is the order we follow in our own readiness checks.

What we see from the data owner’s side

Most companies we talk to assume their support history is a cost center archive. We think it is an asset that has gone unused. The same ticket threads that train your agents contain the real-world language AI developers struggle to find, and the public sets are not licensed for commercial work.

Our approach is built around one principle: you keep your records. Mode is designed so that your company retains ownership and only de-identified copies are shared. Guidance and readiness checks keep all of your company data private until a formal agreement is reached. Nothing leaves your systems on the strength of a conversation.

We say this plainly: earnings depend on the quality and volume of what you share. Our program describes a potential range of $100K to $5M, and where a company lands in that range depends on its data. That is why we start with a fit check rather than a pitch.

If you hold emails, support tickets, chat, documents, CRM history or code repositories, you may be sitting on more than you think.

Fredrik

Who this fits

Our marketplace is aimed at companies with a long operating history and a lot of written or recorded customer interaction. Typical candidates are software firms with support tickets and chat logs, financial services firms with servicing correspondence, and consulting firms with client email and documents.

Owners and COOs usually care about the revenue. General counsel cares about consent, contracts and de-identification. Both questions belong in the same early conversation, so we recommend involving counsel from the first call.

Check whether your data fits

If you want to know whether your records could be licensed, start with our company data licensing fit check. It helps you see which record types you hold and what a buyer would look at first. For a team-wide view across several departments, there is a fit check for teams.

Nordicmartech

Prefer to talk it through? You can book a 20-minute call or request a data licensing consultation. We keep your company data private until a formal agreement is reached.

Questions to ask any buyer

Whoever wants your data, a few questions separate serious buyers from casual ones.

If a buyer cannot answer these clearly, the deal is not ready.

FAQ

What is a call transcript dataset?

It is a collection of written records of phone conversations, usually with speaker labels and metadata, used to train or test language models. It differs from call records, which only log timing and routing details, and from audio sets, which include the recordings.

Can I use a public call center dataset in a commercial product?

Not always. The largest public English transcript set uses a CC BY-NC 4.0 license, which prohibits commercial use, resale and redistribution. Always read the license of any set before training a product on it.

Do I need customer consent to license my call data?

You need a lawful basis for the use, and the FTC treats collecting and using voice data without adequate consent as a violation of the FTC Act. Your general counsel should review your recording notices, state recording laws and customer contracts before any data is shared.

Do I give up ownership if I license my records?

On our platform, no. You keep ownership of your records, and only de-identified copies are shared with a buyer under a signed licensing agreement.

How much could my data earn?

Our program describes potential earnings from $100K to $5M, depending on the quality and volume of the data. It is a range, not a promise, and a fit check is the quickest way to see where your records might sit.

Sources

Recommended