Legal transcription turns hearings, depositions, police interviews, client meetings and legal dictation into text. The standard is exacting: every speaker correctly labelled, every negation preserved, and case names and citations spelled exactly. Automatic speech recognition now produces most first drafts, but where a transcript may become part of the record, the original audio and a human check remain the authority.
Key takeaways
- Legal transcripts are judged on speaker attribution and exact wording, not readability. "I did see him" and "I didn't see him" differ by one word and the whole case.
- Terminology includes Latin phrases, party names, statute references and citation formats such as ECLI:EU:C:2014:317. General models get these wrong in predictable ways.
- Choose verbatim or clean verbatim deliberately. Hesitations can matter in witness testimony and are noise in a dictated letter.
- Keep the audio. ASR drafts with word timestamps make hours of recording searchable, but the recording is the evidence.
- Confidentiality obligations include legal professional privilege, GDPR Article 10 for criminal offence data, and the Law Enforcement Directive for police use.
Where legal transcription is used#
- Court and tribunal hearings. Many courts record proceedings in audio. Transcripts are produced on request for appeals, or routinely in some jurisdictions and case types.
- Depositions and examinations. In the United States, Federal Rule of Civil Procedure 30(b)(3) allows deposition testimony to be recorded by audio, audiovisual or stenographic means. Court reporters and stenographers remain common for the certified record.
- Police and investigative interviews. In England and Wales, PACE Code E governs audio recording of suspect interviews. Transcripts of these recordings are prepared for disclosure and trial.
- Arbitration and internal investigations. Long, multi-party sessions, often in more than one language.
- Legal dictation. Lawyers dictating letters, attendance notes and pleadings. Single speaker, close microphone, dense terminology.
- Client meetings and calls. Increasingly transcribed and summarised for file notes, which raises privilege and confidentiality questions discussed below.
Transcript standards#
Legal users specify the transcript style up front. The choice changes what counts as an error.
| Element | Verbatim | Clean verbatim |
|---|---|---|
| Fillers ("uh", "um") | Kept | Removed |
| False starts, repetitions | Kept: "I, I went, I drove" | "I drove" |
| Non-speech events | [pause], [crosstalk], [inaudible 00:14:32] | Only where they affect meaning |
| Grammar | As spoken | Lightly corrected |
| Typical use | Testimony, interviews | Dictation, meeting notes |
Most modern ASR models are trained on data that removes disfluencies, so their raw output is closer to clean verbatim. Producing a true verbatim transcript usually needs a model or setting that keeps fillers, or human editing against the audio.
Speaker labels are role-based ("THE COURT", "COUNSEL FOR THE CLAIMANT", "WITNESS") rather than "Speaker 1". Diarisation splits the audio by voice; mapping each voice to a role is a separate step, often done once by a human and propagated. In courtrooms with a microphone per position, a multichannel recording makes attribution far more reliable than diarisation on a mixed feed.
Terminology and citations#
Legal speech mixes ordinary language with vocabulary a general model rarely sees.
- Latin and law French: res judicata, mens rea, prima facie, obiter dicta, voir dire. mens rea becomes men's rear.
- Party and place names: every case has its own. A list of parties, witnesses and counsel from the case file is the most useful context a system can get (see rare words and proper nouns).
- Statutory references: spoken "article six, paragraph one, point f" should become "Article 6(1)(f)". This is an inverse text normalisation problem as much as a recognition one.
- Case citations: a UK neutral citation such as [2019] UKSC 41, or a European Case Law Identifier. ECLI identifiers follow a fixed structure: ECLI, a country code, a court code, the year and an ordinal. The Court of Justice's Google Spain judgment is ECLI:EU:C:2014:317; a Dutch Supreme Court decision starts ECLI:NL:HR.
Citations are structured identifiers: their format is known in advance, so a system can be told to expect it. Spoken as E C L I, N L, H R, twenty nineteen, twelve thirty-four, a general model might return easily NL HR 2019 1234; format-aware biasing can recover ECLI:NL:HR:2019:1234.
Accuracy when the transcript is evidence#
A transcript used in proceedings is a claim about what was said. The practical consequences:
- The audio is authoritative. Keep the original recording, with its hash and chain of custody, alongside the transcript. Treat ASR output as a navigable index into it.
- Word timestamps matter. They let a reviewer jump from any word to the audio, and let parties cite a passage by time. See word timestamps and forced alignment.
- Negations and modal verbs are critical errors. "Can" vs "can't", "would" vs "wouldn't". WER counts these as one error each; a reviewer should treat them as the most important ones.
- Marking uncertainty beats guessing. An [inaudible] tag with a timestamp is more useful than a plausible wrong word. Confidence scores can flag passages for review, with the caveat that they are imperfectly calibrated.
- Long recordings. Hearings run for hours. Models must not drift, drop sections or hallucinate in long pauses. See long-form transcription.
Confidentiality and regulation#
Recordings of client meetings are covered by professional secrecy and, in many jurisdictions, legal professional privilege. Sending them to a transcription service is usually compatible with privilege when the service acts on the lawyer's instructions under confidentiality obligations, but firms should check the rules of their bar and jurisdiction and the vendor's contract terms, including whether the vendor retains audio or trains on it.
Under the GDPR, personal data about criminal convictions and offences is governed by Article 10, which restricts processing to official authority or where Union or Member State law authorises it. Police and prosecutors processing interview recordings for law enforcement fall under the Law Enforcement Directive, Directive (EU) 2016/680, rather than the GDPR. Suspects and accused persons in criminal proceedings have a right to interpretation and translation under Directive 2010/64/EU, so multilingual recordings are common.
The EU AI Act lists, in Annex III, AI systems intended to assist a judicial authority in researching and interpreting facts and the law, and in applying the law to facts, as high-risk. Plain transcription is generally not described that way, but a system that summarises evidence for a judge may need closer analysis. See the EU AI Act and speech recognition.
How Reson8 approaches this#
Legal is one of the industries Reson8 serves. Case documents, witness lists and glossaries can be supplied as text through domain adaptation, with up to 1M tokens of context applied at runtime and taking effect in under 60 seconds, so a firm can adapt per matter without collecting audio. Patterns biases the model toward an expected identifier format, which suits case numbers and citation formats. Processing runs on EU GPUs with zero audio retention and no training on customer data; on-premise and VPC deployment are available on enterprise plans.
Common questions#
Can an ASR transcript be used in court?
That depends on the jurisdiction and the court's rules. In most settings, a transcript that will be relied on is checked or certified by a human against the recording. ASR speeds up the draft; it does not replace the verification step.
How should overlapping speech be handled?
Mark it. A [crosstalk] tag with a timestamp is better than a merged sentence attributed to one speaker. Per-speaker microphones reduce the problem at source.
Does sending client recordings to a vendor waive privilege?
Generally not, when the vendor processes on the firm's instructions under confidentiality terms, but the rules vary by jurisdiction. Check your professional body's guidance and the vendor's retention and training terms.
Verbatim or clean verbatim for depositions?
Verbatim. Hesitations, self-corrections and pauses can bear on credibility, and opposing counsel may rely on them.
References#
- Federal Rules of Civil Procedure, Rule 30: Depositions by Oral Examination. Cornell LII.
- Home Office. Police and Criminal Evidence Act 1984 (PACE) Code E: audio recording of interviews with suspects. GOV.UK.
- Council conclusions inviting the introduction of the European Case Law Identifier (ECLI) (2011). Official Journal C 127, 29.4.2011.
- Regulation (EU) 2016/679 (General Data Protection Regulation), Article 10. EUR-Lex.
- Directive (EU) 2016/680 (Law Enforcement Directive). EUR-Lex.
- Directive 2010/64/EU on the right to interpretation and translation in criminal proceedings. EUR-Lex.
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), Annex III. EUR-Lex.