Every barrister and solicitor who has tried general-purpose dictation knows the moment of frustration. You dictate “DPP v Murphy, 2023 IECA 45” and the transcript reads “DPP the Murphy 2023 I see a 45.” You say “res judicata” and get “race you the Carter.” You reference “Ó Domhnaill v Merrick” and the tool produces something that doesn’t resemble either name.

Legal language is uniquely hostile to general dictation tools. It combines proper nouns, formal citation formats, Latin phrases, Irish names with fadas, and domain-specific abbreviations in ways that expose every weakness in consumer speech-to-text.

Why general dictation fails on legal language

General dictation tools — iPhone dictation, Google Voice Typing, even Dragon without legal vocabulary — are trained on everyday language. Their language models predict the next word based on what words typically follow each other in normal speech. This works well for emails and messages. It fails badly for legal language because legal language violates normal language patterns constantly.

Case names are proper nouns in unexpected combinations. “Donoghue v Stevenson” isn’t a pattern that appears in general English. Neither is “Maguire v Ardagh” or “Blascaod Mór Teo v Commissioners of Public Works.” The dictation tool has no prior expectation of these word combinations and falls back on guessing — usually badly.

Neutral citations have a rigid format that looks like gibberish to a general model. The Irish neutral citation system follows a specific pattern: [year] court code number. So “[2024] IEHC 123” means the 123rd judgment of the Irish High Court in 2024. A general dictation tool has no concept of this format. It hears “2024 I E H C 123” and tries to make sense of each element independently, typically producing something like “2024 I see HC 123” or “2024 IAHC 123.”

The Irish court codes are particularly problematic. IEHC (High Court), IECA (Court of Appeal), IESC (Supreme Court), IECC (Circuit Court) — these abbreviations are unfamiliar to models trained on general English. Even Dragon Legal, which knows US and UK court systems, doesn’t natively recognise Irish neutral citations.

Latin terms are ubiquitous in legal writing. Res judicata, obiter dicta, ultra vires, prima facie, inter alia, habeas corpus — these appear in virtually every legal opinion and submission. General dictation tools handle common Latin phrases inconsistently. Less common ones — cestui que trust, quantum meruit, nemo dat quod non habet — are almost always mangled.

Irish names with fadas add another layer. Case names involving Irish-language names — Ó Domhnaill, Ó Laighléis, Ní Chonaill — require the dictation tool to handle both Irish phonetics and the fada diacritics. General tools trained on English don’t attempt fadas and typically produce garbled anglicisations.

The Irish accent compounds every problem

All of these challenges are made worse by the Irish accent factor. General dictation tools are trained primarily on American English, with some accommodation for standard British English. Irish English — with its distinctive vowel sounds, consonant realisations, and prosody — produces higher base error rates even before legal terminology enters the picture.

When an Irish barrister dictates “judicial review in the matter of Ó Domhnaill v Merrick, [2024] IESC 15, the court held that res judicata applied” — that single sentence contains an Irish accent, an Irish name with fadas, a neutral citation with an Irish court code, and a Latin legal term. It’s a stress test that no general dictation tool passes.

The result is that barristers and solicitors who try dictation end up spending as much time correcting citation errors as they saved by dictating. For many, this is the specific frustration that made them give up on dictation entirely.

How legal-trained AI handles citations differently

The difference between general dictation and legal-trained AI dictation isn’t just a bigger vocabulary. It’s a fundamentally different approach to understanding what the speaker is saying.

Context-aware prediction. When a legal AI model hears something that sounds like a case citation, it doesn’t try to match each word independently. It recognises the pattern — a name, “v”, another name, possibly followed by a year and court code — and applies legal context to resolve ambiguous sounds. “DPP v Murphy” is predicted as a unit, not as three separate words.

Citation format recognition. A legal-trained model knows what neutral citations look like. When it hears “[2024] IEHC 123,” it recognises the format and applies the correct structure — square brackets around the year, the correct court code in capitals, the case number. It doesn’t hear “I E H C” as four separate letters to be interpreted individually.

Irish legal vocabulary. A model trained on Irish legal texts knows the Irish court hierarchy, common Irish case names, and Irish legal terminology. It’s encountered “Ó Domhnaill v Merrick” in its training data. It knows that IESC, IECA, IEHC, and IECC are court codes. It doesn’t need to guess.

Latin phrase recognition. Legal AI models treat Latin legal terms as vocabulary items, not as mysterious foreign sounds. Res judicata is in the model’s vocabulary just as “contract” is. The model recognises the phrase in context — after a discussion of prior proceedings, “res judicata” is the predicted term, not “race you the Carter.”

What this means in practice

The practical difference is the gap between dictating and having a usable document.

With a general dictation tool, a barrister who dictates a paragraph containing case citations and Latin terms will spend a significant portion of their time correcting transcription errors. For many practitioners, this tips the balance — dictation ends up slower than just typing the document from scratch.

With a legal-trained AI dictation tool, the same paragraph comes through with the citations correctly formatted, the Latin terms spelled right, and the Irish names rendered with fadas where appropriate. The barrister reviews for substance, not for transcription errors.

This is the difference between dictation that works in theory and dictation that works in practice. The reason dictation never took hold at the Irish Bar wasn’t that barristers didn’t want to dictate — it was that the tools couldn’t handle what barristers actually say. That’s changed.