Industries / Finance and business services / Legal and professional services / Court reporting
Transcripts
Turning the record into a written transcript: its format, its price, who may read it and how long it takes.
7
Milestones achieved, from 1778 to 2026.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1778
A full account required
The City of London required the Old Bailey Proceedings to "contain a true, fair, and perfect narrative of the whole evidence upon the trial of every prisoner", and trial reports at once grew much longer. Even so, the legal historian John Langbein found that in the 1780s they "were still omitting most of what was said at most of the trials they reported."
LessonA rule that demands a full account does not by itself produce one.
Credit
City of London, Committee for City Lands
1907
Shorthand notes for appeals
England's Criminal Appeal Act 1907 said "Shorthand notes shall be taken" at the trial of anyone on indictment who could appeal, with a transcript made for the new Court of Criminal Appeal when its registrar directed. The cost of the notes and the transcripts was paid from money voted by Parliament.
Credit
UK Parliament
1944
One page for every court
Since the Court Reporter Act of 1944, US federal court proceedings must be recorded verbatim. The Judicial Conference set one transcript format that year "to assure that each party is treated equally throughout the country": 25 lines of text a page and about 63 characters a line. The Conference still sets the maximum price; from 1 October 2024 it is $4.40 a page for a transcript delivered in 30 days and $8.70 a page for one delivered in two hours.
Credit
US Congress; Judicial Conference of the United States
2007
Transcripts online after 90 days
In September 2007 the Judicial Conference adopted a policy under which federal court transcripts become available through PACER, the federal courts' paid records website, "after 90 calendar days".
Credit
Judicial Conference of the United States
2024
A recording nobody could transcribe
On 12 November 2024 Ohio's Ninth District Court of Appeals reversed a conviction in State v. Riddlebarger and sent the case back for a new trial. The court reporter had said a transcript of the trial recording "would result in many 'inaudible' lines", and the court found the record too thin to review the appeal. In England, the 2006 High Court transcript of Post Office v Castleton, now among the evidence of the Post Office Horizon IT Inquiry, notes that "due to the poor standard of recording it has not been possible to produce a high quality transcript".
LessonA recording works as a record only if someone can turn it into words.
Credit
Ohio Ninth District Court of Appeals; WordWave (Castleton transcript)
2025
AI writes down witness evidence
The Kerala High Court made Adalat AI's voice-to-text tool compulsory for recording witness depositions in the state's district courts from 1 November 2025, after a pilot in some Ernakulam courts from 1 February. ThePrint reported in October 2025 that over 4,000 courts in nine Indian states had adopted AI, and that early trials in Malayalam "produced transcripts riddled with errors".
Credit
Kerala High Court; Adalat AI
2026
A transcript supplier collapses
The Federal Court of Australia told lawyers that "In March 2026, VIQ's Australian operations entered voluntary administration", and the Supreme Court of Western Australia warned of "disruptions to the provision of transcript". England's courts, by contrast, buy transcription from 8 suppliers under a Ministry of Justice framework that runs from June 2023 to June 2027.
LessonA court that hands its transcripts to a contractor depends on that contractor staying in business.
fedcourt.gov.ausupremecourt.wa.gov.aucontractsfinder.service.gov.uk
Credit
Federal Court of Australia; Supreme Court of Western Australia; UK Ministry of Justice
Watch for
AI transcripts for Crown Courts
The Ministry of Justice is testing whether its own AI tool, Justice Transcribe, "could meet required accuracy standards while reducing transcription time and costs" on Crown Court hearings, and says victims will get free transcripts of judges' sentencing remarks from spring 2027.
Achieved whenHMCTS publishes accuracy results against its contracted human transcribers and the free transcripts are in routine use.
Claimed by
UK Ministry of Justice, April 2026
Adalat AI's accuracy figure
Adalat AI's chief legal officer told ThePrint that the tool "was 95 per cent accurate all the time."
Achieved whenan independent measurement, or a published Kerala High Court review of the courts' monthly reports, gives an accuracy figure.
Claimed by
Adalat AI, October 2025
Open
01
An open transcript file format
A documented, open file format for court transcripts that keeps page and line numbers (lawyers cite transcripts by page:line), speaker labels, question and answer, colloquy (talk among the judge and lawyers) and exhibit marks, with converters from the plain text and PDF that reporters' software puts out.
The briefCredit · unclaimed02
Check transcripts against audio
A tool that lines a finished transcript up with its recording and with a separate speech-to-text pass, and flags passages where the words or the speaker probably differ, for a person to listen to.
The briefCredit · unclaimed
Open 01 · the brief
An open transcript file format
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A documented, open file format for court transcripts that keeps page and line numbers (lawyers cite transcripts by page:line), speaker labels, question and answer, colloquy (talk among the judge and lawyers) and exhibit marks, with converters from the plain text and PDF that reporters' software puts out.
Why it is still open
The OASIS LegalXML Legal Transcripts committee was set up to define "an XML compliant syntax for representing legal transcript documents". It is listed as completed with only a document model and no approved standard. US federal policy lets reporters deliver realtime files "in ASCII format, or any other format requested by the ordering party". Akoma Ntoso, an open standard for legal documents, has elements for questions, answers and line ends, but I found no profile of it for court transcripts.
Built with what exists
JSON Schema or an Akoma Ntoso profile; pdftotext and pdfplumber to read transcript PDFs; public transcripts such as the daily transcripts of UK public inquiries and US federal transcripts in the free RECAP archive.
First step
Mark up by hand one page from each of three layouts, such as a US federal transcript, a US state deposition and a UK inquiry transcript.
Done looks like
At least 50 transcripts from at least 3 sources converted into the format and back with every page:line position kept, and at least 98% of elements given the right type on a hand-checked sample.
Open 02 · the brief
Check transcripts against audio
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A tool that lines a finished transcript up with its recording and with a separate speech-to-text pass, and flags passages where the words or the speaker probably differ, for a person to listen to.
Why it is still open
In 1983 the Federal Judicial Center compared transcripts with tapes by hand, using panels of judges and lawyers. A 2006 MIT paper used alignment "to discover and correct errors in the manual transcription", but on lectures, not courts. A 2022 paper warns that "transcripts can technically be 'aligned' with any audio", so alignment alone cannot show that a transcript is right. I found no checker of this kind tested on public hearings with official transcripts.
Built with what exists
a speech-to-text model such as ElevenLabs Scribe v3 (see the Note "Fewer word errors") or an open-weight model; pyannote to label who spoke when; hearing videos and transcripts that the Post Office Horizon IT Inquiry publishes.
First step
Plant 200 changes in five inquiry transcripts and run a first version.
Done looks like
With 200 deliberate changes to words and speaker labels planted in real transcripts, at least 90% of them flagged, with at most 5 flags per hour of unchanged audio, and a listener checking a sample of the flags on unchanged audio.