Deposition and Court Transcripts as Legal Language Data

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Why testimony is missing from open legal corpora

Public legal datasets are built from text that courts and agencies publish, and testimony is rarely published. Pile of Law gathers court opinions, filings, agency publications, statutes, regulations and casebooks [1], and MultiLegalPile likewise compiles public legal text at multilingual scale [2]. Public benchmarks such as LegalBench test reasoning through hand-crafted tasks rather than examination transcripts [8].

The reason is procedural. Under the Federal Rules of Civil Procedure, depositions are generally not filed with the court unless they are used in a motion or at trial, so most transcripts live only with the reporting firm, the noticing party and counsel who ordered copies. Hearing and trial transcripts reach the docket more often, but many sit behind reporter fee periods or sealing. A model trained only on opinions learns how judges write about testimony, not how lawyers elicit it.

For the broader corpus map, see the text datasets hub for LLM training and proprietary text that web crawls do not contain.

What a usable deposition transcript record contains

A training-grade transcript is a sequence of attributed turns, not a flat text file. Court reporters deliver in formats such as ASCII text with page:line numbering, PTX or other e-transcript files, condensed transcripts and PDF, and each carries structure that is easy to destroy during conversion. The fields below are what an SFT, summarization or evaluation pipeline actually consumes.

Illustrative example: invented to show structure; it does not describe an available dataset.

{
  "transcript_id": "T-000412",
  "proceeding_type": "deposition",
  "jurisdiction": "US federal, civil",
  "case_type": "product liability",
  "witness_role": "corporate designee (Rule 30(b)(6))",
  "examining_party": "plaintiff",
  "certified": true,
  "errata_applied": true,
  "turns": [
    {"turn": 118, "page_line": "47:12", "speaker_role": "Q", "speaker": "ATTORNEY_1",
     "text": "Did your team review the field failure reports before the March release?"},
    {"turn": 119, "page_line": "47:15", "speaker_role": "COLLOQUY", "speaker": "ATTORNEY_2",
     "text": "Objection, form."},
    {"turn": 120, "page_line": "47:16", "speaker_role": "A", "speaker": "WITNESS_1",
     "text": "We reviewed a summary, not the individual reports."}
  ],
  "exhibits_referenced": [{"exhibit_no": "EX-14", "description_redacted": true, "produced": false}],
  "confidentiality_designation": "none",
  "deid_method": "rule-based + NER replacement, sample audited",
  "rights_basis": "reporter license + party consent"
}

Three structural points matter most. Separate Q, A and colloquy (objections, instructions not to answer, off-the-record notes), because colloquy pollutes Q/A pairs if it is merged into answers. Keep page:line anchors so summaries and eval items can cite the record the way litigators do. Record whether the transcript is the certified final with the witness's errata applied, or a rough draft, which can differ materially.

The rights layers behind a testimony transcript

Licensing a transcript means clearing several overlapping interests, and the reporter's sign-off alone rarely covers all of them. Court reporting firms typically control transcript copies through engagement terms and per-copy pricing, and some assert copyright in their formatting and indexing. The parties who ordered and paid for the transcript may have contract rights or expectations, and the witness's testimony is personal speech that often includes health, employment and financial details.

Protective orders are the hard constraint. Rule 26(c) lets a court, for good cause, forbid or limit disclosure of discovery material, and standard stipulated orders let parties designate deposition pages "Confidential" or "Attorneys' Eyes Only," usually within a set window after the transcript is delivered. Material under such a designation, or sealed by the court, should be excluded unless the order itself permits the use. Expect designations to be page-level, so a single transcript can be partly usable.

Recent case law makes upstream rights for legal AI a live issue. As of October 2026, the Third Circuit has affirmed partial summary judgment for Thomson Reuters against ROSS over Westlaw headnotes used to build a non-generative legal research tool [6]. That ruling concerned headnotes, not testimony, but it shows that courts will examine how legal training material was obtained. For drafting the grant itself, see writing the AI training rights grant, and for firms holding client material, client data held by service providers.

De-identifying witnesses, parties and third parties

Testimony needs more aggressive de-identification than filings, because it is unscripted and names people who never appear in the caption. Rule 5.2 already requires filers to redact Social Security and taxpayer numbers to the last four digits, birth dates to the year, minors' names to initials and financial account numbers to the last four digits [3], but deposition answers also contain employers, addresses, diagnoses, co-workers and family members. Court-text anonymization research annotates quasi-identifiers and confidential attributes alongside names, because they can also enable re-identification [4].

Medical testimony in personal injury or malpractice cases may contain protected health information from records produced in discovery. Where the source is a HIPAA covered entity or its records, de-identification should follow Safe Harbor or Expert Determination [5]; see what a HIPAA-compliant label must mean. Treat models as possible leak paths, since training data can be extracted from deployed language models by querying them [7].

Practical rules that hold up in review:

• Replace names with consistent role tokens (WITNESS_1, ATTORNEY_2) so dialogue structure survives.
• Generalize dates, places and employers rather than deleting them, to keep the reasoning intact.
• Strip exhibit contents unless separately cleared; keep only exhibit numbers and redacted descriptions.
• Remove case captions and docket numbers that let a reader look up the unredacted record.

Matching transcripts to SFT, summarization and evaluation

Each application needs a different slice of the same corpus, so specify the use before the volume. For supervised fine-tuning on examination, you want clean Q/A turns from certified finals with colloquy labeled and witness role recorded. For summarization, you want whole transcripts with page:line anchors and, ideally, existing attorney-written deposition summaries as targets.

Illustrative example: invented to show structure; it does not describe an available dataset.

For evaluation, split by matter, not by page, or the model will have seen the same witness on both sides of the split. Pair testimony with speech transcript normalization guidance if you also license audio or video depositions.

Diligence questions before you license testimony

Ask for a rights ledger at transcript level before you review samples. It should state the source (reporting firm, law firm or corporate legal department), the proceeding type, whether a protective order exists and what it designates, whether the matter is sealed, whether the witness or parties consented, and the de-identification method with its audit result.

Then check the text itself: the share of certified finals versus rough drafts, whether errata were applied, how colloquy is tagged, and how exhibits are handled. These checks decide whether the corpus trains a deposition assistant or a model that confuses objections with answers. For the contract-focused side of legal AI data, see legal AI training data from real contract work.

How SourceX approaches deposition transcript requests

SourceX sources operational datasets from US companies on request and manages the commercial process, including licensing and ongoing purchases. Transcripts are not held in stock, and a request does not guarantee a match. Buyers describe the data they need; SourceX looks for US businesses that hold it, and every release is approved by the supplying company.

Each dataset is rights-reviewed for ownership and consents and delivered under a license defining records, uses, term and delivery. Personal details are removed or replaced before delivery, the method is recorded and a sample is checked, though no method is perfect. For background on the supply side, read whether court reporting firms can sell data to AI companies and what AI teams build with court reporting data, or describe your testimony data need.

Request deposition transcript data for legal AI

Describe the proceeding types, structure and uses you need, and SourceX will look for US businesses that hold matching transcripts. Nothing is contracted until a supplier agrees, and pricing and allowed uses are set in the license. Start a buyer request at SourceX.