Industries / Finance and business services / Translation services
Interpreting
Interpreters put speech from one language into another as it is spoken, at conferences, on the telephone and over video links, and software now tries to do the same.
7
Milestones achieved, from 1927 to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1927
Interpreting through headphones
Delegates at International Labour Organization (ILO) meetings in Geneva found consecutive interpreting, in which the speaker pauses while the interpreter repeats the speech in another language, too slow. Simultaneous interpreting, in which the interpreter speaks at the same time as the speaker and listeners hear the version they choose through headphones, was first used at the International Labour Conference on 4 June 1927, with equipment that became known as the IBM Hushaphone Filene-Findlay system. The Boston businessman Edward Filene, an employers' representative, asked for a "telephone translation" and helped pay for its development, and Gordon Findlay of IBM designed a multi-channel system, by the account of the interpreter and scholar Ruth Morris.
Credit
Edward Filene; Gordon Findlay; IBM; the ILO
1945
Simultaneous interpreting at Nuremberg
The International Military Tribunal at Nuremberg in 1945–46 ran on IBM equipment lent free of charge, and most of its judges had no language in common. Practically none of the 36 interpreters had used the method before. Speakers were held to 60 words a minute, a monitor listened to each interpreter's output against the original, reviewers checked the shorthand record in each language, and whole teams changed every 90 minutes. Morris writes that without simultaneous interpreting the trial could not have been held.
LessonThe new method worked once speed limits, live monitoring and fixed shifts were built around the people doing it.
Credit
Colonel Léon Dostert, who tested and recruited the interpreters; Alfred Steer, who ran the language division; IBM
1947
The UN makes it permanent
The United Nations at first interpreted its debates consecutively, then ran both methods side by side after Dostert equipped two rooms with booths in 1946. On 15 November 1947 the General Assembly, in resolution 152 (II), decided that "simultaneous interpretation be adopted as a permanent service", to be used instead of or alongside consecutive interpreting as debates required.
LessonThe UN trusted the new method only after running it next to the old one for more than a year.
Credit
the UN General Assembly; Léon Dostert
1973
Interpreters by telephone
Australia's federal government set up a free Emergency Telephone Interpreter Service for people who did not speak English. It took 11,349 calls in its first year. Its successor, TIS National, run by the Department of Home Affairs, calls it the world's first telephone interpreting service and says it now delivers more than one million immediate phone services a year, with more than 2,700 interpreters and translators in over 150 languages.
Credit
the Australian Government
2018
Human work credited to AI
At a forum in Shanghai on 20 September 2018, a screen showed what was presented as machine interpreting by the Chinese speech company iFlytek. The interpreter Bell Wang wrote that the Chinese text on screen was a transcript of his and a colleague's own interpreting: the software misheard "Davos Forum" as "Devil's Forum" in English, yet the Chinese came out right, as theirs had. iFlytek's chief executive said the tool in use was transcription, not translation.
LessonA live demonstration of machine interpreting proves nothing unless the audience knows whether people are working behind it.
Credit
Bell Wang, who made the account public; reported by Sixth Tone
2022
Microphone rules after injuries
When Canada's Parliament sat by video during the COVID-19 pandemic, interpreters at the federal Translation Bureau filed more health and safety reports, citing headaches, earaches and fatigue from poor sound. A study the Bureau commissioned from the National Research Council found that the sound reaching interpreters was worse than what participants connected directly to Zoom heard. From 7 February 2022 the Bureau has refused to interpret for anyone in a Government of Canada meeting who does not use an external unidirectional microphone, and the interpreters' union filed a complaint under the Canada Labour Code on 31 January 2022.
LessonIn remote interpreting, the sound depends on every speaker's microphone, so the rules have to bind the speakers as well as the interpreters.
Credit
Translation Bureau, Public Services and Procurement Canada; National Research Council of Canada; Canadian Association of Professional Employees
2025
Translation inside earbuds
Apple released Live Translation for its AirPods in September 2025, in beta, speaking a translation into the listener's ear during face-to-face conversations in English, French, German, Portuguese and Spanish. In July 2026 the UK consumer group Which? tested it with four bilingual speakers, half the time with background noise, and found it generally accurate, but it mangled names ("British Airways" became "Briti shareway"), translated some Spanish word for word, and lagged on long sentences and in noise.
LessonDelay, more than wrong words, is what made conversation through the earbuds feel unnatural.
Credit
Apple; tested by Which?
Watch for
Live translation in any headphones
Google says a beta in its Translate app speaks real-time translations into any pair of headphones, using Gemini's speech-to-speech translation, which turns speech in one language directly into speech in another and, Google says, keeps each speaker's tone, emphasis and cadence. It launched on Android in the US, Mexico and India with more than 70 languages, with iOS and more countries promised for 2026.
Achieved whenan outside tester, such as a consumer group or university, publishes accuracy and delay results across several languages, as Which? did for Apple's feature.
Claimed by
Google, December 2025
Open
01
A repeated interpreting scorecard
A public scorecard that plays the same recorded European Parliament speeches through the live translation tools people can buy, including the earbud and phone features above, and compares their output with the official human interpretation of the same speeches, scored for meaning errors and delay, and rerun every six months so the trend shows.
The briefCredit · unclaimed02
A remote-interpreting sound bench
A repeatable bench test that sends the same recorded speech and test tones through each common video and remote-interpreting platform, with typical speaker setups (laptop microphone, headset, phone), and measures what arrives at the interpreter's end: which frequencies survive, how loud the peaks get and how much the sound is compressed.
The briefCredit · unclaimed
Open 01 · the brief
A repeated interpreting scorecard
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public scorecard that plays the same recorded European Parliament speeches through the live translation tools people can buy, including the earbud and phone features above, and compares their output with the official human interpretation of the same speeches, scored for meaning errors and delay, and rerun every six months so the trend shows.
Why it is still open
Single comparisons now exist. Tomasz Korybski and colleagues at the University of Warsaw ran two machine interpreting services beside EU-accredited and student interpreters at a live Polish–English event in June 2025: the accredited interpreters made the fewest errors and the machines the most, and more of the machines' errors distorted the meaning. Dongpeng Pan and Kilian Seeber (WMT, October 2026) compared four machine systems with professional interpreters on simulated conference speeches into five languages: the machines' content scores fell within or above the professionals' range, but only one commercial system came close to their timing, and the slowest lagged more than ten times as far behind the speaker as the professionals did. Each is a one-time test, and neither covered the earbud and phone features.
Built with what exists
The ESIC corpus of Dominik Macháček and colleagues (2021), 10 hours of English European Parliament speeches with their interpretation into Czech and German; European Parliament plenary recordings with their interpretation channels; Apple Live Translation and Google Translate's live beta; Pan and Seeber's reference-free scoring, calibrated against expert ratings; a speech-to-text model such as ElevenLabs Scribe v3 to transcribe both the machine and the human output (see the Note "Fewer word errors"); two professional conference interpreters per language to grade blind.
First step
Take 10 ESIC speeches, play them through two tools, and have one Czech interpreter mark meaning errors in both the machine output and the human channel.
Done looks like
100 speeches into at least three languages, each run through at least three tools; meaning errors per 100 words and average delay in seconds for every tool and for the human channel, with the graders' agreement; the speeches, transcripts and scores published, and the same test rerun six months later.
Open 02 · the brief
A remote-interpreting sound bench
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A repeatable bench test that sends the same recorded speech and test tones through each common video and remote-interpreting platform, with typical speaker setups (laptop microphone, headset, phone), and measures what arrives at the interpreter's end: which frequencies survive, how loud the peaks get and how much the sound is compressed.
Why it is still open
Kilian Seeber's laboratory at the University of Geneva ran an experiment with more than 80 participants and found that the narrower sound typical of online platforms, compared with the full range AIIC and ISO recommend, raised their workload and lowered the quality of their output as judged blind. Canada's Translation Bureau had the National Research Council measure sound on Parliament Hill. Neither source gives a platform-by-platform comparison, and this research found none that is public and repeatable.
Built with what exists
A calibrated USB measurement microphone; Room EQ Wizard, free software that plays sweep tones and plots frequency response; a small loudspeaker standing in for the speaker's mouth; paid accounts on Zoom, Microsoft Teams and two remote-interpreting platforms; recorded speech from a public corpus; Python to compare runs. All loud-peak tests are recorded to file, with nobody listening on headphones.
First step
Set up two accounts on one platform, send a sweep tone from a laptop microphone to a second computer, and plot what arrives with Room EQ Wizard.
Done looks like
For at least five platforms and three microphone setups, the measured frequency range, peak level and loudness range at the receiving end, each measured three times and repeatable within 1 dB; the method, recordings and results published so anyone can rerun them after a platform update.