Industries / Finance and business services / Legal and professional services / Court reporting
Live text
Turning the record into text on a screen as the words are spoken, for the judge and lawyers, or for deaf and hard-of-hearing people in the room.
4
Milestones achieved, from 1991 to 2023.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1991
Live transcripts for lawyers
LiveNote, a service built by Graham Smith-Bernal, put the reporter's text on lawyers' and judges' screens during a hearing, and was ready for clients by 1991, MoneyWeek reports. The firm was later sold for $70 million. By 1999 a rival had launched a system that it hoped would "challenge the market dominance of the LiveNote product", Law.com reported.
Credit
Graham Smith-Bernal
1995
Live text for deaf jurors
On 7 December 1995 Camille Jones became, in NCRA's words, "the first person with hearing loss to sit on a jury in Los Angeles County Superior Court". She followed the trial through CART (communication access realtime translation, a writer's live text on a screen) provided by Jacquie Gutierrez, and the court's policy has been in place since 1999. US federal rules under the Americans with Disabilities Act list "real-time computer-aided transcription services" among the aids that state and local governments must offer.
LessonNCRA's guidance says one person should not act as both the official reporter and the CART captioner. /CART-in-the-Court-Setting-a-Standard).
Credit
Jacquie Gutierrez; Camille Jones; Los Angeles County Superior Court
2011
Streamed hearings, daily transcripts
The Leveson Inquiry into the culture and ethics of the British press streamed its hearings live online from November 2011 and said "Transcripts will be posted on the Internet daily." Merrill Legal Solutions produced the transcripts.
Credit
Leveson Inquiry; Merrill Legal Solutions
2023
AI live text in India
On 21 February 2023 the Supreme Court of India began showing an AI live transcription of a Constitution Bench hearing on screens in the courtroom, a first according to ThePrint. Chief Justice D.Y. Chandrachud said staff would clean up the errors by the evening, giving "a permanent record of arguments".
Credit
Chief Justice D.Y. Chandrachud; Supreme Court of India
Watch for
Live captions at the EPO
The European Patent Office offers automated live captions, on request, in oral proceedings held by video, for "equitable access for parties with hearing impairments", and says it "cannot guarantee accuracy".
Achieved whensomeone outside the EPO measures the captions' accuracy or delay, or reports their routine use.
Claimed by
European Patent Office, April 2026
Real-time AI in tribunals
The Ministry of Justice's AI unit says Judicial Transcribe "helps courts and tribunals capture proceedings in real time" and, after pilots, "has been rolled out nationally across the Immigration and Asylum Chamber". Its algorithmic transparency record is marked as still in progress.
Achieved whenthe transparency record and an independent evaluation are published.
Claimed by
UK Ministry of Justice, September 2026
AI realtime with human editors
Opus 2, which supplies live transcripts for court hearings and arbitrations, announced AI realtime transcription checked by "experienced editors", in testing with clients it did not name.
Achieved whena named court, inquiry or arbitration uses it routinely, or independent accuracy and delay figures are published.
Claimed by
Opus 2, June 2026
Open
01
Timing AI live captions
A public score of AI live captions on recorded hearings played back as if they were live. It measures the delay in seconds from a word being spoken to its appearing on screen and to its final version, and the accuracy, above all on names and numbers, against the official transcript.
The briefCredit · unclaimed02
Case word lists from filings
A tool that reads a case's public court filings and produces the names, places and terms a reporter must prepare before a hearing. It outputs them as a steno job dictionary (a case-only word list loaded into the reporter's software) and as a keyterm list for speech-to-text, and measures how many proper nouns then come out right.
The briefCredit · unclaimed
Open 01 · the brief
Timing AI live captions
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public score of AI live captions on recorded hearings played back as if they were live. It measures the delay in seconds from a word being spoken to its appearing on screen and to its final version, and the accuracy, above all on names and numbers, against the official transcript.
Why it is still open
The closest study measures television: a September 2026 paper compares TV captions "with up to 7-12 seconds delay" with AI captions averaging about two seconds. The EPO, the Ministry of Justice and Adalat AI publish no measured delay. A 2025 study of UK Supreme Court hearings measured legal terms on recordings after the fact, not live. I found no delay test on courtroom audio.
Built with what exists
streaming speech-to-text services such as Deepgram Nova-3 and ElevenLabs Scribe v3 Live (see the Note "Fewer word errors"); forced alignment of the official transcript to the audio, to find when each word was really spoken; archived hearing video with daily transcripts, such as the Leveson Inquiry's.
First step
Pick one Leveson Inquiry hearing day, check that its video is available, and align it with its transcript.
Done looks like
A public table for at least three systems on at least 10 hours from at least five hearings, giving the median and 95th-percentile delay in seconds, the word error rate and the error rates on names and numbers, with code that anyone can rerun.
Open 02 · the brief
Case word lists from filings
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A tool that reads a case's public court filings and produces the names, places and terms a reporter must prepare before a hearing. It outputs them as a steno job dictionary (a case-only word list loaded into the reporter's software) and as a keyterm list for speech-to-text, and measures how many proper nouns then come out right.
Why it is still open
CasePrepd (version 1.0.35, May 2026) pulls vocabulary from case documents into Word, PDF or text files, but makes no steno dictionary, reports no accuracy and does not read public court records. The 2025 UK Supreme Court study built its term list from judgments and transcripts after the hearings. A March 2026 paper says "there is no standardized benchmark" for custom vocabulary in speech recognition.
Built with what exists
CourtListener's RECAP archive, whose documents include their extracted text; use only documents already in the archive, because fetching new ones from PACER costs money. Also spaCy to find names and places; Deepgram's keyterm prompting, which takes up to 500 tokens per request; and Plover's dictionary format.
First step
Take five cases that have both briefs and an argument transcript, pull the names and places from the briefs, and count how many of the transcript's proper nouns the list covers.
Done looks like
For at least 20 cases with public filings and a later public transcript: the share of the transcript's names and places that the list caught, the share of list entries that appear in the transcript, and the name error rate of speech-to-text with and without the list.