Captioning legal department and law firm training video: Latin legal vocabulary, citation format failures, CLE obligations, and ADA requirements for in-house legal teams

The legal training video landscape

Legal training video is produced in two distinct institutional contexts: law firms producing content for their own attorneys and staff, and corporate legal departments producing training for in-house lawyers and non-lawyer staff. Both contexts create distinct training needs, and the vocabulary challenges in each differ in emphasis rather than kind.

Law firm training video

At large law firms (AmLaw 200 and above), the volume of training video is substantial. Practice group training, litigation skills development, new associate orientation, ethics and professional responsibility compliance, diversity and inclusion training, and client-specific matter briefings all generate video at scale. The AmLaw 100 firms have dedicated professional development and L&D teams; mid-size firms (50–200 attorneys) typically have an HR director or office manager who runs training without dedicated L&D infrastructure. Both types produce video. The difference is in the sophistication of the LMS and the captioning workflow — or its absence.

CLE is the defining feature of the law firm training landscape that distinguishes it from other verticals. Almost every U.S. jurisdiction requires licensed attorneys to complete a specified number of CLE credit hours per year (typically 12–15 hours, varying by state). Firms increasingly produce their own accredited CLE programmes as a cost-reduction measure and as a way to deliver training that is specific to the firm’s practice areas and client base. When that CLE content is delivered in video format — as is now the standard for most bar-approved on-demand CLE — captioning becomes a compliance obligation, not a feature.

In-house legal department training video

In-house legal departments at companies with 50 or more employees face a different training profile. Their video library typically includes: legal hold and e-discovery procedure training, contract management platform onboarding, compliance training (FCPA, antitrust, data privacy, employment law), legal ops software training (Ironclad, Clio, SimpleLegal, Mitratech), and department-specific policy updates. The vocabulary in this content is legal-technical but also heavily company-specific: client names, contract management platform terminology, internal matter codes, and jurisdiction-specific regulatory abbreviations (GDPR, CCPA, FCPA, HIPAA in the context of healthcare legal departments).

The common thread between law firm and in-house legal training is that the vocabulary appearing in the video is vocabulary a viewer may use to perform legal work. A Deaf attorney who watches a CLE video on pleading standards needs the citations right. A legal ops analyst who watches a contract management onboarding video needs the platform terminology right. This use case — caption as reference material — creates an accuracy standard for legal content that is higher in practice than the standard that applies to general L&D training, even though the formal legal threshold (WCAG 2.1 AA at 99%+) is the same across verticals.

Why standard LMS auto-captions fail legal content harder than other verticals

Every major LMS with built-in auto-captioning — Absorb, TalentLMS, Docebo, Cornerstone, Kaltura, Canvas Studio — uses a general-purpose ASR engine as its captioning backend. These engines train on large corpora of general English speech. They perform adequately on business English, conversational language, and commonly occurring technical vocabulary. They perform poorly on vocabulary that is highly domain-specific and low-frequency in standard speech corpora.

Legal vocabulary falls into three categories of ASR failure that compound on each other. Category one: Latin terms that appear essentially nowhere in speech training corpora. Category two: citation formats and abbreviations that require verbatim accuracy. Category three: case name proper nouns that are the names of private individuals and companies not otherwise present in general vocabulary. A CLE video on appellate procedure may contain all three failure categories in the same sentence: “Under Twombly, a complaint must plead sufficient facts to state a claim for relief that is plausible on its face — 550 U.S. 544 (2007) — and the Iqbal formulation extended this to require the court to draw on its judicial experience and common sense.” This single sentence contains two case names with near-zero ASR coverage, a reporter citation, and implicitly invokes the res judicata and stare decisis context a presenter would likely be discussing nearby. The aggregate WER of the rest of the paragraph might be excellent. The words that matter are wrong. See the caption quality error rate calculator for how aggregate WER can mask vocabulary-specific failure.

ADA Title I obligations for law firms and legal departments

The starting point for any legal team evaluating its captioning obligations is ADA Title I, which prohibits employment discrimination against qualified individuals with disabilities by covered employers. Title I requires covered employers to provide reasonable accommodations to qualified individuals with disabilities, absent undue hardship.

Coverage threshold

ADA Title I applies to employers with 15 or more employees. In the law firm context, this threshold is calculated on the basis of full-time and part-time employees, and includes paralegals, legal assistants, administrative staff, and non-attorney professionals alongside attorneys. Virtually all firms above the very smallest boutique practices (solo practitioners and two- or three-attorney partnerships) meet or exceed this threshold. The AmLaw 100, AmLaw 200, and most mid-size regional firms are clearly covered. See the U.S. caption compliance matrix for the ADA Title I coverage framework and how it interacts with state employment discrimination statutes, which often apply at lower thresholds than the federal 15-employee floor.

How Title I applies to training video

ADA Title I’s reasonable accommodation requirement covers all employment-related training: new hire orientation, professional development, compliance training, and job-specific skills training. When a law firm delivers training in video format and a Deaf or hard-of-hearing employee requests accessible captions, the firm is obligated to engage in the interactive process and provide an effective accommodation. The question of what constitutes an “effective” accommodation in the captioning context is the crux of the legal analysis.

Courts and the EEOC have interpreted “effective communication” in the context of training content to mean that the accommodation enables the individual to participate in the training on equal terms with non-disabled colleagues. Auto-generated captions that render “McDonnell Douglas” as “McDonald Douglas” and “stare decisis” as “stare decides” do not provide effective communication for an attorney who needs those terms right in order to perform the professional development function the training is designed to achieve. An accommodation that is formally present but functionally inadequate does not satisfy the reasonable accommodation standard.

The practical implication is that an L&D team or HR department at a law firm that uses LMS auto-captions as a default for all video and calls it “captioned” without checking accuracy on legal vocabulary has not discharged its ADA Title I obligation. It has produced captions that will fail on the vocabulary that matters most to a legal professional viewer.

CLE as employment training

CLE is the area where law firms most commonly fail to recognise the ADA Title I implication. When a firm requires associates to complete a specific number of CLE hours per year as a condition of employment, professional advancement, or partnership track eligibility, that CLE is employment training. It is not optional educational content that the firm happens to make available; it is training that has a performance and employment consequence attached to non-completion.

If the firm delivers that CLE in video format — as firm-produced CLE is almost universally delivered today — and a Deaf or hard-of-hearing associate requests captions, the ADA Title I interactive process applies. The accommodation standard is effective communication, which in the CLE context means captions accurate enough to convey the legal content of the programme in a form the associate can use. Captions that render Latin terms, citations, and case names incorrectly do not meet that standard, regardless of what aggregate word accuracy the vendor reports. See building a caption compliance programme for L&D for the framework for managing accommodation requests within a training programme.

In-house legal departments

In-house legal departments are covered by their employing company’s ADA Title I obligations, not by any separate legal department framework. The company’s HR function manages accommodation requests, but the in-house legal team is responsible for the accessibility of training content it produces and delivers to legal department employees. When a legal operations team produces a contract management platform training series or a data privacy compliance video and a Deaf legal ops analyst enrolls, the analysis is the same as for any other employment training — the standard is effective communication, and the specific vocabulary failures of legal content apply in full.

ADA Title III and CLE provider obligations

ADA Title III applies to places of public accommodation, which the statute defines to include private schools and educational institutions. When a law firm submits CLE content for bar approval and makes it available to external bar members — not just to the firm’s own employees — the firm may be acting as a CLE provider subject to Title III, in addition to the Title I obligations that apply to the firm’s own employees.

The public accommodation analysis

The Title III public accommodation analysis for CLE providers is not settled, and courts have reached different conclusions about whether an online-only CLE provider with no physical location constitutes a “place of public accommodation” under the statute’s text. The ADA was enacted before widespread online education, and courts have split on whether virtual educational offerings trigger Title III independently of any physical location. The Department of Justice has taken the position in recent guidance that online-only services of covered entities are subject to Title III; the case law remains mixed.

The practical risk calculus for a law firm producing CLE for external distribution is: if the content is accessible (captioned accurately), the Title III question is moot. The accessibility improvement that eliminates the Title I risk for the firm’s own employees eliminates the Title III exposure as well. Building the legal captioning workflow correctly solves both problems simultaneously.

State bar CLE accessibility requirements

State bars are adding their own accessibility requirements for approved CLE providers independently of the ADA framework, and these requirements vary significantly by jurisdiction. California’s State Bar has issued guidance encouraging accessible CLE format, and the California CLE regulations contemplate that providers should accommodate bar members with disabilities. New York’s CLE Board has addressed accessibility in its provider standards. The Florida Bar and Texas State Bar have both referenced accessibility in CLE provider guidelines.

The specific requirements are in flux across jurisdictions, and a law firm’s compliance counsel or CLE programme coordinator should verify current requirements in any state where the firm produces or submits approved CLE content. What is consistent across jurisdictions is the direction of movement: state bars are adding accessibility requirements to CLE provider approval criteria, and firms that establish a captioning workflow now are positioned to maintain provider status as those requirements solidify. The compliance training captions reference page covers the regulatory landscape for captioning obligations across multiple frameworks.

Latin legal vocabulary: the core failure category

The vocabulary failure that most distinguishes legal training from all other L&D content is Latin. Hundreds of legal concepts are expressed in Latin phrases that entered English legal usage centuries ago and have remained in active use. The problem for ASR is straightforward: these terms appear with near-zero frequency in the speech corpora on which ASR models train.

Why Latin fails ASR structurally

ASR models learn to decode speech by training on very large corpora of audio paired with transcripts. The models learn the relationship between phoneme sequences and word forms by exposure — they have seen (and heard) the word “reasonable” millions of times in their training data, so they decode it correctly. They have seen “mens rea” approximately never. The consequence is not just a wrong word guess; the model is making a phoneme-to-word mapping without any relevant prior exposure, and the phoneme sequences in Latin are not typical English sequences. Latin vowel qualities (the long /ae/ in stare decisis, the /oe/ in habeas corpus, the French-derived phonology of voir dire) are systematically different from English vowel patterns, and the model has no basis for correct decoding.

A secondary factor is that Latin legal terms are often said aloud in court and in legal education with anglicised pronunciations that are themselves non-standard: the legal pronunciation of voir dire is “vore dire” or “vore dye-er” in American courtrooms, not the French /vwɑr diʁ/. The anglicised form is not in the ASR training corpus either, because it does not appear in podcast audio, news broadcasts, or YouTube video at any meaningful frequency. The model receives a phoneme sequence it has essentially never seen before and produces whatever word or phrase matches most closely — which in practice is a non-word substitution or a semantically unrelated phrase.

Term-by-term failure analysis

The following analysis covers the Latin terms most commonly encountered in CLE video and legal department training. The “typical ASR output” column represents observed failure patterns across major commercial ASR engines on unglossed legal content; the specific output will vary by vendor and by the audio characteristics of the recording, but the failure rate at the word level is consistently high across all general-purpose engines.

Vocabulary accuracy versus aggregate accuracy in legal content

The critical insight from the table above is that Latin term failures are not evenly distributed across the transcript. A 30-minute CLE video on appellate procedure might contain 200 instances of Latin terms drawn from this vocabulary, interspersed among 5,000 words of common English. The common English may be captioned at 98% accuracy. The Latin terms may be captioned at 30–50% accuracy at the word level. The aggregate document WER — the number most auto-caption quality dashboards report — will look acceptable (perhaps 95–97%) while the vocabulary that is most important to a legal professional viewer is substantially wrong.

This is the same structural problem described for technical vocabulary in engineering training (where SDK symbols and API method names fail), medical training (where drug names and anatomical terms fail), and cybersecurity training (where MITRE ATT&CK technique IDs fail). In each case, the vocabulary most important to the viewer is the vocabulary ASR is least equipped to handle, and aggregate metrics hide the failure. See why 99% caption accuracy matters and the engineering glossary approach for the underlying framework.

Citation format failures: the verbatim accuracy problem

Legal citation formats present a different failure mode from Latin vocabulary. The challenge is not that ASR cannot decode the spoken words — the words “twenty-eight USC section thirteen thirty-one” are perfectly ordinary English — but that the caption renders them in a format that is not the conventional citation format, and the viewer needs the format right.

Why format accuracy matters in legal content

In other L&D verticals, the question of caption accuracy is framed in terms of comprehension: does the viewer understand the content the speaker intended? In legal training, there is a secondary use case that is equally important: the viewer may use the information in the caption as a reference for later work. A Deaf paralegal watching a legal holds procedure training and seeing the citation “18 CFR 1.7” in the caption may copy that citation into a legal hold notice or research memo. If the citation is wrong or incomplete, the professional error propagates downstream from the training content.

This is a use case no other L&D vertical shares at the same scale. The accuracy standard for legal training captions is therefore not just the WCAG 2.1 AA 99%+ word accuracy standard (though that applies) but a reference-document standard: the caption must be accurate enough that a professional can rely on it for subsequent work. Citation formats are the sharpest edge of this problem.

Common citation format failures by citation type

U.S. Code citations (28 U.S.C. § 1331): Spoken as “twenty-eight U.S.C. section thirteen thirty-one.” LMS auto-captions typically render this as either the spoken form (“twenty-eight U.S.C. section thirteen thirty-one”), a partial numeric form (“28 U.S.C. section 1331”), or an abbreviated form (“28 USC 1331”). None of these forms preserves the section symbol (§), which is the conventional citation element used in legal writing. The presenter said “section,” but the conventional abbreviated form is “§” — ASR transcribes the spoken word, not the conventional symbol. A viewer who wants the standard-form citation must reconstruct it from what the caption says, which introduces a step that defeats the reference-material use case.

Federal Rules of Civil Procedure (Fed. R. Civ. P. 12(b)(6)): Spoken in different ways by different presenters: “Federal Rules of Civil Procedure twelve-B-six,” “Rule twelve-B-six,” “FRCP twelve-B-six.” The conventional citation form (Fed. R. Civ. P. 12(b)(6)) includes parenthetical notation for subsections that spoken language typically omits. ASR renders whatever was said, which is usually not the standard citation form. The (b)(6) notation — which specifies the particular ground for dismissal and is legally significant — may be rendered as “B6” or “B six” rather than “(b)(6).”

Code of Federal Regulations (18 C.F.R. § 157.10): Similar to U.S. Code citations: spoken as “eighteen C.F.R. section one fifty-seven point ten,” rendered variably by ASR. The “C.F.R.” abbreviation may be decoded as “CFR” or “C.F.R.” depending on whether the presenter spells out each letter or runs them together, and the period in subsection numbers (“157.10”) may be lost or rendered as a sentence-end period.

ERISA and other statutory citation systems: ERISA § 502(a)(1)(B) is a common citation in employment and benefits CLE. Spoken as “ERISA section five-oh-two-A-one-B,” the caption typically produces “ERISA section 502 A 1 B” rather than the standard “ERISA § 502(a)(1)(B)” form. The parenthetical structure, lower-case subsection letters, and section symbol are all lost.

Reporter citations (550 U.S. 544 (2007)): The citation to Twombly is “550 U.S. 544 (2007).” Spoken as “five-fifty U.S. five forty-four two thousand seven,” ASR typically renders the numbers correctly but the format inconsistently. The “(2007)” parenthetical is often lost or rendered as “2007” without parentheses. For older cases where the citation form uses “(19__)” — for example, Miranda v. Arizona, 384 U.S. 436 (1966) — the year is likely to be decoded correctly as “1966,” but the overall citation format will not match convention.

The correct response: citation glossary entries and format templates

The solution for citation format accuracy is not to expect ASR to produce correct citation format from spoken audio — it cannot, because the conventional format includes symbols and punctuation that have no spoken equivalent. The correct approach is a two-step workflow. First, the caption is produced with best-effort ASR output. Second, the transcript is reviewed against known citation standards before publication, with a reviewer checking every instance of a citation pattern and correcting the format to convention.

This is qualitatively different from vocabulary glossary correction, which pre-processes the ASR output before delivery. Citation format correction is a post-processing step that requires a human reviewer with citation knowledge. For CLE content that is intended to be used as a reference by attorneys, this step is not optional if the goal is effective communication. The caption QA methodology post covers the general QA framework; legal content adds a citation-specific check that should be part of every review before publication.

Case name proper noun failures

The names of parties in reported appellate decisions are proper nouns with near-zero representation in general ASR training corpora. The reason is simple: case names are the names of individuals and companies involved in specific lawsuits, and they appear in legal text, law review articles, briefs, and court opinions — none of which constitute a significant fraction of the audio corpora on which ASR models train. The model has essentially no exposure to these names in audio form, and so it produces phoneme-level substitutions that may or may not produce recognisable output.

Case name failure patterns

The pattern analysis reveals that the failure rate for case names correlates with three factors: (1) whether the surname has entered common English vocabulary through other means; (2) whether the phoneme sequence is typical English; and (3) whether the word appears in a frequently cited cultural context outside of law. The following table covers the most commonly cited U.S. cases in CLE and legal department training content:

The McDonnell Douglas problem as a case study

McDonnell Douglas Corp. v. Green, 411 U.S. 792 (1973), is the foundational case for the Title VII employment discrimination burden-shifting framework. It appears in virtually every CLE video on employment law and every employment law practice group training. The McDonnell Douglas burden-shifting framework is referred to by name dozens of times in a single training module. And it fails consistently, because the ASR model has vastly more exposure to “McDonald” (McDonald’s fast food) than to “McDonnell” (the aircraft manufacturer that merged with Douglas in 1967 to form McDonnell Douglas, itself later merged with Boeing). The one-letter difference in spelling corresponds to a different pronunciation (/məkˈdɔnəl/ versus /məkˈdɔnăld/) but ASR decodes based on phoneme similarity and prior frequency, and “McDonald” dominates both.

An employment law CLE video in which every mention of the McDonnell Douglas framework is captioned as “McDonald Douglas” is a captioning failure on the most legally significant term in the module. It is also the kind of failure that a Deaf attorney who relies on captions as their access channel will notice immediately — and that a non-Deaf reviewer might not catch in a casual review, because “McDonald Douglas” is superficially close enough to pass a quick scan.

Practice area case name glossaries

The practical implication of the case name failure analysis is that every practice area that is the subject of CLE or training content should have a dedicated glossary of the most commonly cited cases in that area. Employment law: McDonnell Douglas, Faragher, Ellerth, Vance, Harris, Oncale, Bostock. Antitrust: Leegin, Trinko, Twombly, Actavis. Evidence: Daubert, Joiner, Kumho, Frye. Civil procedure: Twombly, Iqbal, Pennoyer, Celotex, Anderson, Matsushita. Constitutional law: Marbury, McCulloch, Lochner, Carolene Products, Chevron, Mead. The glossary for each practice area is manageable in scope (20–40 case names per area) and transforms the caption accuracy on case names from near-zero to reliable. See customer glossary architecture for AI captions for the implementation framework.

Legal procedure and document vocabulary

Beyond Latin terms and case names, legal training video contains a vocabulary layer of legal procedure terms and document designations that are not Latin but are still domain-specific and variably handled by ASR.

Common legal procedure terms and their ASR handling

Court and tribunal name formats

Court names present a specific formatting challenge. Attorneys speaking in CLE video routinely abbreviate court names: “the Second Circuit,” “the Southern District,” “SDNY,” “N.D. Cal.” The spoken forms are typically decoded correctly by ASR (the phoneme sequences are ordinary English), but the format of the decoded output — whether the abbreviation is presented as “SDNY,” “S.D.N.Y.,” or “S-D-N-Y” — will vary and will not match the conventional citation form (S.D.N.Y.) without glossary guidance.

For CLE content that cites specific circuit court opinions, the numbered circuit name is important: the First Circuit handles different geographic territory than the Ninth Circuit, and an attorney studying circuit split analysis needs to know which circuit issued which opinion. “The First Circuit held” and “the Ninth Circuit held” are both decoded correctly by ASR. The problem arises when the presenter switches between numbered citation form (“1st Cir.”) and spoken form (“the First Circuit”), and the caption needs to be consistent.

E-discovery vocabulary

E-discovery training has become a staple of in-house legal department L&D, driven by the prevalence of e-discovery obligations in commercial litigation and the growing complexity of the data landscape. E-discovery vocabulary has its own ASR failure profile, distinct from the Latin/case name failures of CLE content.

E-discovery terms and their ASR handling

ESI (electronically stored information): Spoken as either “E.S.I.” (spelled out) or “e-si” (as a word). When spelled out, ASR typically renders it as “ESI” or “E.S.I.” When spoken as a word, it may be rendered as “easy” or “Esi.” The frequency of the spelled-out form in training data is low; the spoken-as-word form even lower. The term appears in virtually every e-discovery training module.

TAR (technology-assisted review): Almost universally decoded as “tar” (the common English word) rather than as an acronym for technology-assisted review. Context rarely rescues this: “We used tar to review the document set” is grammatically plausible to the ASR model.

Custodian (e-discovery): The e-discovery term for a person whose data is collected. “Custodian” is a common English word and is decoded correctly; the challenge is that the e-discovery meaning differs from the common meaning (a janitor or building manager), and a viewer who relies on the caption alone will understand the word without ambiguity only if they already know the e-discovery context.

De-NISTing: The process of filtering out known system files (identified by NIST hash values) to reduce the document review population. Spoken as “de-NISTing” or “de-Nisting.” ASR typically renders this as “de-nesting” or fails to produce a recognisable output. The term is too domain-specific to appear in training data at useful frequency.

FRCP 37 (Rule 37 sanctions for discovery failures): Spoken as “FRCP thirty-seven” or “Rule thirty-seven.” The “FRCP” abbreviation is decoded as “F.R.C.P.” or “FRCP” inconsistently; the number is usually correct.

Clawback (FRCP 26(b)(5)(B)): The rule governing the inadvertent production of privileged documents and the mechanism for retrieving them. “Clawback” is decoded correctly as a compound word. The Rule citation has the format challenges described above.

Predictive coding: The machine-learning approach to document review. Decoded correctly; the compound is common enough in technology discussion to appear in training data.

The e-discovery vocabulary glossary for legal department training should include: ESI, TAR, de-NISTing, NIST, custodian (with context), litigation hold, legal hold, clawback, FRCP 26/37, metadata, deduplication, near-duplicate analysis, email threading, and the platform-specific terms for whatever e-discovery software the organisation uses (Relativity, Reveal, Everlaw, Nextpoint, Casepoint). See glossary architecture for how to structure platform-specific vocabulary alongside domain vocabulary.

Building the legal glossary: four vocabulary tiers

The legal vocabulary challenge is solvable with a structured glossary, but the glossary for legal content has a different architecture than the glossary for engineering or medical training. Legal vocabulary organises into four distinct tiers, each with different update cadence and different maintenance ownership.

Tier 1: Latin vocabulary

Latin legal terms are stable — they have been in use for centuries and do not change. The Latin tier of the glossary is a one-time build that requires no ongoing maintenance beyond occasional addition when a new training module covers a Latin term not previously in the library. The tier should include all terms that appear in any practice area the firm or department trains on, with phonetic pronunciation hints where the spoken form departs from an obvious reading (particularly for terms like voir dire, certiorari, and in limine where the anglicised legal pronunciation is specific and non-obvious).

The Latin tier glossary entry format should include: the Latin phrase, the conventional anglicised pronunciation as phonetically spelled, and the legal meaning as a disambiguation anchor. Example entry: voir dire — vore-dire — jury examination process; motion to examine witness on qualifications. The pronunciation hint helps the ASR system match the anglicised spoken form to the correct glossary entry.

Tier 2: Citation formats and abbreviations

Citation format abbreviations — U.S.C., C.F.R., F.R.C.P., F.R.E., U.S., F.2d, F.3d, F.Supp.2d, S.Ct., L.Ed. — appear with high frequency in any CLE or legal department training content and require consistent format in the output. The citation tier of the glossary should include the spoken forms (how presenters say these abbreviations) mapped to the conventional citation format. Example: spoken “U.S.C.” (spelled out) → output “U.S.C.”; spoken “USC” (run together) → output “U.S.C.”. The section symbol (§) needs special handling, since it has no spoken equivalent; the reviewer-applied post-processing step (described in the citation format section above) handles this.

Tier 3: Case names by practice area

The case name tier is the most labour-intensive to build but the most impactful for legal professionals who rely on CLE captions as reference material. The tier should be organised by practice area: employment law case names, antitrust case names, evidence case names, civil procedure case names, constitutional law case names, and so on. Each practice area should include 20–40 of the most commonly cited cases in CLE and training content for that area.

The update cadence for the case name tier is event-driven: when a significant new Supreme Court or circuit court opinion is issued and CLE content about it is created, the case name tier is updated within 24–48 hours. In a year with major Supreme Court terms, this may mean 5–10 new case names per year. See caption glossary maintenance workflow for the event-driven update process and the 48-hour SLA standard for new term additions.

Tier 4: Organisation-specific and platform-specific vocabulary

The fourth tier covers vocabulary specific to the law firm or legal department: client names that are frequently referenced in training content (where training modules use real client matter examples or hypotheticals built around client industry), internal platform names (the contract management platform, the matter management system, the billing platform), and regulatory abbreviations specific to the organisation’s practice areas (specific agencies, specific regulations). This tier has the same maintenance cadence as the Tier 3 case name tier: event-driven, with new terms added when a new training module covers new vocabulary. See the caption feedback loop post for how to build a sustainable maintenance workflow that captures new vocabulary before it causes a recurring failure.

Glossary build timeline

Building all four tiers of the legal glossary from scratch is a 2–3 week project for a single legal L&D professional with access to the existing training library and a list of the practice areas covered. The Latin tier (Tier 1) can be built in a single work session of 2–4 hours by working through the most common Latin terms in a standard legal dictionary or bar exam preparation resource. The citation tier (Tier 2) requires 1–2 hours and produces a list of 20–30 abbreviation forms. The case name tier (Tier 3) requires a review of the existing training library to identify which cases are referenced in each module; for a firm with 5–10 active practice areas and 20–50 training videos, this produces 100–200 case names. The Tier 4 organisational vocabulary is specific to the firm and requires a review of the training content itself to identify the terms used.

The time investment in the glossary build should be understood as a one-time cost that amortises across every video in the training library and every future module the firm produces. Without the glossary, every CLE module requires a manual caption review session of approximately 30–60 minutes per video before publication (the correction-labour model described in the hidden half-FTE post). With the glossary, the review is a final-pass citation format check, which takes 10–15 minutes. For a firm that produces 10 CLE videos per year, the glossary build pays back in reduced review time within the first three videos.

Multi-speaker formats in legal training

Legal training video frequently uses multi-speaker formats: moot court simulations, mock trial exercises, panel discussions among attorneys, and Q&A sessions between a presenter and a moderator. Multi-speaker audio is harder to caption accurately than single-speaker audio for two reasons: speaker diarisation (identifying who is speaking at any given moment) is an additional processing challenge, and the informal, overlapping speech patterns of panel discussions produce audio quality issues that degrade ASR accuracy on top of the vocabulary challenges described above.

Moot court and mock trial formats

Moot court training video is particularly challenging because it contains dense, rapid exchanges between attorneys and judges with heavy use of Latin terms (attorneys arguing before a court say certiorari, stare decisis, amicus curiae, and citation strings at high density) in an audio environment that includes multiple speakers interrupting each other and speaking quickly under pressure. These conditions produce the worst-case ASR performance profile for legal vocabulary.

The recommendation for moot court and mock trial training content is to separate the glossary preparation and the human review investment: ensure the Latin tier and case name tier of the glossary are fully built before the first moot court module is captioned, and build the human review time for multi-speaker verification into the production workflow. See the caption QA methodology post for the review workflow.

Panel discussions and CLE Q&A

Many law firm CLE productions use a panel format, with two to four attorneys discussing a legal topic interactively. The informal speech register of panel discussions produces more hesitations, false starts, and overlapping speech than prepared presentations — all of which degrade ASR accuracy. Combined with legal vocabulary density, panel discussions should be budgeted for longer caption review times than prepared-presentation videos of the same length. The human transcript review pass for a 30-minute CLE panel discussion will typically take 20–30 minutes rather than the 10–15 minutes for a prepared single-speaker presentation.

Procurement considerations for legal training video

When a law firm or legal department evaluates a caption vendor for CLE and training content, the standard procurement process — covered in the captioning RFP playbook — needs to be augmented with legal-specific requirements. The vendor’s general accuracy claims (99% on soft-skills studio content) are not relevant to the question of how they will perform on legal vocabulary. The specific questions to add to a legal L&D captioning RFP are:

• Latin legal vocabulary handling: Does the vendor support custom glossary entries for Latin phrases? How does the glossary system handle multi-word Latin terms (e.g., mens rea, voir dire, stare decisis) where each word must be treated as part of a fixed phrase? Can the vendor demonstrate accuracy on a sample file containing Latin legal vocabulary before engagement?
• Accuracy on legal content specifically: What is the vendor’s benchmark accuracy on legal training content as distinct from general business content? If the vendor cannot distinguish these, that is informative: they are treating legal training as equivalent to HR onboarding, which it is not.
• Citation post-processing workflow: Does the vendor have a human review step that checks citation formats? How is this priced? What turnaround time applies to citation-reviewed content?
• Case name glossary: Does the vendor offer pre-built legal case name glossaries, or does the client build the glossary from scratch? If from scratch, what format does the glossary take and how long is the initial build?
• Confidentiality: For law firms handling client-specific training content, the confidentiality of the captioned content is important. Does the vendor sign an NDA or BAA? Is training content processed on shared infrastructure or isolated? Legal L&D content may be more sensitive than general corporate training content because it may contain client names, matter strategy, and confidential legal analysis.

The vendor pilot for legal training content should include a diagnostic sample from the firm’s actual CLE or training library: a 5–10 minute clip from an employment law module (which will contain McDonnell Douglas, Iqbal/Twombly, and EEOC/ADA citations) and a clip from an appellate procedure module (which will contain Latin terms at high density). The go/no-go threshold for the pilot should be set before the sample is submitted, not after reviewing the results. See caption vendor pilot programme design for the full pilot methodology and the pre-committed threshold framework, and how to evaluate a vendor’s RFP response for reading vendor accuracy claims critically before the pilot begins.

Contract provisions for legal content captioning

The caption vendor contract for legal training content should include provisions beyond the standard accuracy SLA. Specific additions to consider:

• Legal vocabulary accuracy SLA: A separate accuracy measurement for vocabulary-specific accuracy on a diagnostic content sample, in addition to the aggregate WER SLA. The legal vocabulary accuracy SLA should cover at minimum Latin terms and case names included in the glossary.
• Glossary update SLA for new case names: When a significant new opinion is issued that will be covered in upcoming CLE content, a 24–48 hour SLA for adding the case name to the glossary and making it available for new captioning jobs.
• Citation format review as a contractual deliverable: If the contract includes human review, specify that citation format checking is part of the review scope, not just accuracy QA.
• Confidentiality and data handling: The NDA or confidentiality provisions should specifically address how legal training content is processed and stored, whether it is used for model training, and how long it is retained. See the vendor SLA and contract review checklist for the full framework.

The third-party CLE compliance training dimension

Many law firms and legal departments use third-party CLE providers in addition to (or instead of) producing their own content: PLI (Practising Law Institute), LexisNexis CLE, Westlaw Precision (formerly West LegalEd Center), ABA continuing education, state bar programme offerings. When a firm purchases access to third-party CLE video for its attorneys and a Deaf or hard-of-hearing attorney requests accessible captioning, the analysis is more complex than when the firm controls the content.

Third-party content accessibility obligations

Under ADA Title I, the firm’s reasonable accommodation obligation extends to the provision of accessible training regardless of whether the firm produced the training content. If the firm assigns an associate to complete a PLI CLE module as part of their development plan and the PLI content is not accessible, the firm’s obligation to provide the accommodation does not disappear because the content was produced by a third party. The firm must either: (a) source accessible captioned content from the third-party provider; (b) arrange for captioning of the third-party content before the associate accesses it; or (c) source alternative accessible content that covers the same material.

This analysis mirrors the third-party compliance training captioning obligation covered in detail in the third-party compliance training captioning post. The legal-specific wrinkle is that CLE credit requirements may constrain the firm’s ability to substitute alternative content: if the state bar has approved specific PLI courses for specialisation credit, the firm cannot simply substitute different content that covers the same topics but lacks the bar approval.

Engaging CLE providers on accessibility

Large CLE providers (PLI, LexisNexis, Thomson Reuters/Westlaw) have accessibility programmes of varying quality, and most are improving in response to litigation and regulatory pressure. When a firm or legal department has a Deaf attorney who requires CLE access, the first step should be to contact the CLE provider’s accessibility team directly and determine what accessible format options exist. Most major providers can deliver captioned content or human-reviewed transcripts for specific programmes on request, sometimes with advance notice requirements of 2–4 weeks for human-reviewed content.

Where the provider cannot deliver accessible content within the required timeframe, the firm’s accommodation obligation may require arranging independent captioning of the content, which requires a contract with the provider that permits captioning for accessibility purposes. The accessibility coordinator playbook covers the general third-party content accommodation workflow; legal L&D should have the legal department review the contract terms for any CLE platform agreement to ensure accessibility accommodation rights are preserved.

GlossCap and legal department captioning

GlossCap’s glossary-biased decoding architecture is particularly well-suited to legal training content because the problem is fundamentally a vocabulary problem: ASR vocabulary coverage of Latin terms, case names, and citation abbreviations is insufficient, and the solution is a glossary that provides coverage before ASR decoding begins.

For law firms and legal departments setting up a GlossCap workflow for CLE and training content, the implementation path is:

1. Build the legal glossary using the four-tier structure described above. The Tier 1 (Latin) and Tier 2 (citation abbreviations) portions can be imported from a template covering the most common legal vocabulary; the Tier 3 (case names) and Tier 4 (organisational vocabulary) portions are customised to the firm’s practice areas and the specific modules in the training library. GlossCap applies the full glossary at decode time, not as a post-processing text substitution, so the accuracy benefit applies to the initial ASR output rather than requiring a separate correction pass.
2. Set the citation format post-processing step as a required workflow step for any CLE or legal training content before publication. This is a human reviewer step that takes 10–15 minutes per module and ensures citation formats are in conventional form. GlossCap’s edit UI supports this step natively, allowing the reviewer to search for citation patterns and correct format without re-reviewing the full transcript.
3. Maintain the case name glossary tier on a 24–48 hour SLA whenever significant new opinions are issued in a practice area the firm trains on. For most firms, this means a quarterly review of recent Supreme Court terms and relevant circuit court opinions, plus real-time updates during active term periods.

GlossCap delivers WCAG 2.1 AA captions with DCMP per-cue scoring documentation, SRT and VTT output for Absorb, Docebo, TalentLMS, Canvas, Cornerstone, and all major LMS platforms, and glossary-biased decoding that addresses the Latin, case name, citation, and organisational vocabulary tiers. For CLE content that requires submission to a state bar with an accessibility standard, we can provide the DCMP methodology documentation required for compliance purposes.

If your firm or legal department is building a CLE captioning workflow from scratch — or is dealing with an existing accommodation request for accessible CLE content — start with the GlossCap widget preview and our pricing tiers. The Team plan (30 hrs/mo, Notion/Confluence/Docs glossary sync) is the right starting point for most law firm L&D teams; the Org plan (unlimited hours, SSO, custom glossary model) is appropriate for AmLaw 100 firms or legal departments with large training libraries.

Frequently asked questions

Does ADA Title I really apply to CLE that the firm requires its associates to complete?

Yes. ADA Title I covers all employment-related training provided by a covered employer. When a law firm requires associates to complete CLE hours as a condition of employment, professional advancement, or partnership track eligibility, that CLE is employment training subject to the ADA Title I reasonable accommodation standard. The fact that CLE is bar-required (i.e., the underlying obligation comes from the state bar rather than the firm) does not remove the employer’s ADA obligation for the specific delivery mechanism the employer uses or requires. If the firm delivers CLE in video format and a Deaf associate requests captions that provide effective communication, the ADA Title I interactive process applies.

Our firm has 12 attorneys but 18 total employees (including paralegals, legal assistants, and administrative staff). Are we covered by ADA Title I?

Yes. ADA Title I coverage is based on total employees, not attorneys. Paralegals, legal assistants, legal secretaries, administrative professionals, and non-attorney staff all count toward the 15-employee threshold. A firm with 12 attorneys and 6 non-attorney staff employees has 18 total employees and is clearly covered. Many small boutique firms with fewer than 10 attorneys are covered by ADA Title I because their non-attorney staff brings total headcount above the threshold. Check with your HR counsel or employment attorney for the specific count, including part-time employees who are counted at the appropriate fraction.

What is the accuracy standard that applies to CLE captions under the ADA?

The ADA Title I standard is “effective communication” as a reasonable accommodation, which does not specify a percentage accuracy threshold directly. In practice, EEOC guidance and enforcement have treated captions as failing the effective communication standard when they contain errors that prevent the viewer from accessing the substantive content of the training. For legal training, where the vocabulary includes Latin terms, case names, and citation formats that are all professionally significant, the effective communication bar is high — a caption that renders voir dire as “vore dare” and Iqbal as “Ick-ball” throughout a 90-minute employment law CLE module does not provide effective communication to an attorney who needs those terms right. The WCAG 2.1 AA 99%+ standard under DCMP per-cue methodology is the technical benchmark most compliance programmes use, and it is a reasonable proxy for the effective communication threshold in legal content. See why 99% caption accuracy matters for the relationship between the technical standard and the ADA accommodation analysis.

We use YouTube auto-captions for CLE modules hosted on our YouTube channel. Is that compliant?

No, for two reasons. First, YouTube auto-captions on legal content produce the Latin, case name, and citation errors described throughout this post — errors that are not corrected by default in the YouTube platform. Second, YouTube’s auto-captions are not reviewed or QA’d by anyone; the output reflects the ASR engine’s performance on your content, which for legal vocabulary will be significantly below the 99% threshold. If you are using YouTube as your CLE delivery platform, you can upload a corrected SRT or VTT subtitle file to the video, which replaces the auto-generated captions with your reviewed transcript. GlossCap delivers SRT and VTT files ready for YouTube upload. See the WCAG 2.1 AA captions reference page for what the WCAG 2.1 AA 99% standard technically requires.

How do we handle Latin pronunciation in the glossary so the ASR decodes it correctly?

The practical approach is to include both the conventional English spelling (voir dire) and a phonetic representation of how the term is spoken in American legal contexts (“vore-dire” or “vore-dye-er”). Glossary-biased decoding systems use the phonetic form to guide the ASR output; the output is then the conventional spelling. For a term like certiorari, the phonetic form is “sir-shee-oh-rar-ee” in most American courtroom usage. The Latin tier of the glossary should include both forms for any term where the anglicised pronunciation departs significantly from a naive reading of the Latin. When in doubt about pronunciation, the legal pronunciation guide in the Bluebook (for terms used in legal writing and argument) or a legal dictionary with pronunciation notes is the reference source.

Our state bar does not have explicit captioning requirements for CLE. Are we still obligated to caption CLE content for our attorneys?

Yes, if an attorney at your firm with a disability requests captions as an accommodation. The ADA Title I obligation is independent of state bar CLE captioning requirements. State bar requirements (where they exist) address the CLE provider’s obligation to external bar members; ADA Title I addresses your firm’s obligation to your own employees. If the state bar does not require accessible CLE, that does not affect your firm’s ADA Title I obligation to your employees. The two obligations operate independently and both may require accessible content from different angles. Additionally, your state employment discrimination law may have a lower employee threshold than ADA Title I (15 employees). California FEHA applies at 5 employees; New York Human Rights Law at 4; New Jersey LAD has no minimum. See the U.S. caption compliance matrix for state-level employment discrimination law coverage thresholds.

How should we handle captioning for a multi-day CLE programme where different presenters speak on different days, including some external speakers?

Multi-day CLE programmes with multiple presenters are handled with the same glossary applied across all sessions, with supplemental entries for any presenter-specific vocabulary (a guest speaker who is an expert in a niche area may use highly specific vocabulary not covered in the base glossary). For external speakers, the key preparation step is to request their speaker materials and bibliography in advance: the case names, statutory citations, and technical terms in their slides and handouts give you the vocabulary to add to the glossary before the session. If the programme is delivered live (with real-time captioning rather than post-production ASR), the advance materials are even more important — CART providers use them to prepare their stenography before the event. See the CART captioning post for the advance materials process in live settings.