BRENNEN

Industries / Government, education and community / Government / Emergency services

Taking calls and dispatch

Answering emergency calls, working out what is wrong and where, sending the right crews, and keeping in touch with them by radio and data link.

7

Milestones achieved, from 1985 to 2018.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1985

    CPR coached by phone

    Dispatchers in King County, Washington, talked callers through cardiopulmonary resuscitation (CPR), pressing on the chest of a person whose heart has stopped, while crews were on the way. M. S. Eisenberg and colleagues described the method in 1985. A study of 7,265 cardiac arrests in the county from 1983 to 2000 found that 25.7% of patients got CPR from a bystander coached by a dispatcher, and that these patients had higher adjusted odds of leaving hospital alive than those who got no bystander CPR (odds ratio 1.45). Centers elsewhere now audit this coaching against American Heart Association (AHA) measures, as S. R. Dowker and colleagues did on 65 calls in a 2022 study.

    ebi.ac.ukebi.ac.ukebi.ac.uk

    Credit

    M. S. Eisenberg, T. D. Rea and colleagues, King County Emergency Medical Services

  2. 1991

    One number across Europe

    On 29 July 1991 the Council of the European Communities decided that every member state must make 112 work on fixed and mobile phone networks as a single European emergency number, alongside national numbers, by 31 December 1992, with delays allowed to 1996 where changes were hard. The decision notes that the European Conference of Postal and Telecommunications Administrations (CEPT) had recommended 112 in 1976, but only a very small number of countries had followed.

    eur-lex.europa.eu

    Credit

    Council of the European Communities; CEPT

  3. 1992

    London's dispatch computer fails

    On 26 October 1992 the London Ambulance Service moved all of London onto a new computer-aided dispatch (CAD) system, software that logs calls and picks ambulances. The official inquiry found the software "not complete, not properly tuned, and not fully tested", staff not fully trained, and no paper backup in the control room. Response times at times became unacceptable on 26 and 27 October, and on 4 November a minor programming error crashed the system because the switch to the backup had not been tested. Press reports linked up to 20 deaths to the breakdown; the inquiry said no coroner's court had found that a late ambulance caused a death.

    LessonA dispatch system must be tested at full load, with a working fallback and trained staff, before it goes live.

    dcs.gla.ac.uk

    Credit

    inquiry team led by Don Page for the South West Thames Regional Health Authority

  4. 2001

    Radios fail in the towers

    On 11 September 2001 the New York fire department's radios worked poorly inside the World Trade Center towers. The 9/11 Commission found that the main tactical channel was overwhelmed, most units that entered the South Tower did not use a repeater channel installed to boost signals there, and some firefighters were on the wrong channel or had no radio. Most Port Authority police officers in the North Tower lacked access to the channel on which their evacuation order was given. The Commission concluded that the radio failure contributed to the firefighters' deaths in the North Tower but was not the main cause.

    LessonCrews need radios that work inside tall buildings and a plan for which units use which channel.

    9-11commission.gov

    Credit

    National Commission on Terrorist Attacks Upon the United States

  5. 2004

    Scripted questions find more arrests

    The Medical Priority Dispatch System gives call-takers a fixed script of questions that sorts medical calls by urgency. Staff of the London Ambulance Service reported in 2004 that after it was adopted, the number of calls correctly identified as cardiac arrest rose by 200%. In Melbourne, Australia, a 2006 study matched three months of calls to patient records and found the script caught 76.7% of cardiac arrests, with 99.2% of other calls correctly ruled out.

    LessonA script raised detection but still missed about one arrest in four in Melbourne's count.

    ebi.ac.ukebi.ac.uk

    Credit

    A. Heward and colleagues, London Ambulance Service; J. Flynn and colleagues, Melbourne

  6. 2015

    Volunteers sent by text

    In a randomized trial in Stockholm from April 2012 to December 2013, a phone-positioning system found volunteers trained in CPR within 500 m of a suspected cardiac arrest and sent them there, alongside ambulance, fire and police crews. Bystanders started CPR before crews arrived in 62% of arrests where volunteers were sent, against 48% where they were not. More than 9,800 volunteers signed up during the study.

    ebi.ac.uk

    Credit

    M. Ringh, L. Svensson and colleagues, in the New England Journal of Medicine

  7. 2018

    Cars call 112 themselves

    Since 31 March 2018 every new type of passenger car and light van approved in the European Union must carry eCall. After a serious crash it opens a voice line to the emergency center, either automatically through sensors in the car or at the press of a button, and sends a minimum set of data that includes the car's position.

    eur-lex.europa.eu

    Credit

    European Parliament and Council, Regulation (EU) 2015/758

Watch for

  1. AI interpreters on 911 calls

    Motorola Solutions says its Interpreter Agent, built into its dispatch software, identifies a caller's language within seconds and gives two-way, voice-to-voice translation with a transcript. The company says waiting for a human interpreter can delay a response by more than 70 seconds; it gives no accuracy figures for the translations.

    Achieved whenan outside group publishes, for test calls in several languages, how often key details such as the address and the patient's breathing come through correctly, and how much time is saved against a human interpreter line.

    motorolasolutions.comcriticalcomms.com.au

    Claimed by

    Motorola Solutions, June 2026

  2. Text and video to 112

    The European Accessibility Act requires emergency centers to answer 112 using the same means the caller used: voice and real-time text, which appears letter by letter as it is typed, together, and video where the caller sends it. The duty applies from 28 June 2025, but member states may put it off until 28 June 2027.

    Achieved whenthe European Commission or national regulators report that emergency centers in every member state answer 112 calls made in real-time text.

    eur-lex.europa.eu

    Claimed by

    the European Union, in Directive (EU) 2019/882, with obligations starting June 2025

Open

  1. 01

    Emergency calls through AI interpreters

    A public set of scripted emergency calls in at least 10 languages, spoken by native speakers and by synthetic voices, with background noise and panicked delivery. Each script carries a fixed list of key facts: address, callback number, what happened, how many people are hurt, and whether the patient is awake and breathing. Run every call through several AI interpreters and score how many key facts reach the dispatcher's language correctly and how fast. Use only test systems or recordings, never a live emergency line.

    The brief
    Credit · unclaimed
  2. 02

    Time phone CPR from audio

    Software that listens to a recorded emergency call and marks the moments the AHA's telecommunicator CPR measures use: when the call is answered, when the call-taker recognizes a cardiac arrest, and when the first chest compression is instructed. It also checks whether the call-taker coached depth, rate and letting the chest rise. Test it on simulated calls in which an actor kneels at a training manikin and a trainee call-taker reads a standard CPR script over a line that reaches only the test room.

    The brief
    Credit · unclaimed

Open 01 · the brief

Emergency calls through AI interpreters

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public set of scripted emergency calls in at least 10 languages, spoken by native speakers and by synthetic voices, with background noise and panicked delivery. Each script carries a fixed list of key facts: address, callback number, what happened, how many people are hurt, and whether the patient is awake and breathing. Run every call through several AI interpreters and score how many key facts reach the dispatcher's language correctly and how fast. Use only test systems or recordings, never a live emergency line.

Why it is still open

Vendors sell AI translation for 911 calls without published accuracy figures. The only spoken 911 translation study found, Carnegie Mellon's NineOneOne project (2008), worked on Spanish calls only, and its classifier recognized address-giving turns in 39% to 63% of cases, depending on the label set. A 2026 paper by Sara Court and colleagues is a case study of a text-to-911 translation service, not a test of spoken calls, and a 2025 German benchmark of emergency speech tested transcription only. The papers at the first workshop on language technology for crisis response, held in November 2026, test translation of written public alerts and pandemic health text, not spoken calls.

aclanthology.orgarxiv.orgpmc.ncbi.nlm.nih.govaclanthology.org

Built with what exists

speech-to-text models such as the open-weight Whisper or ElevenLabs Scribe v3 (see the Note "Fewer word errors"); open translation models and general AI models; text-to-speech voices; public noise recordings; the NineOneOne paper's labels for call content as a starting list of key facts.

  1. First step

    Write 30 call scripts in English with six key facts each, have them translated and recorded in Spanish and one other language, and run them through one open pipeline.

  2. Done looks like

    300 calls in 10 languages, each run through at least three interpreters, with the share of key facts rendered correctly and the median seconds to the first translated fact published per language and per interpreter, and the calls, scripts and scores released.

Open 02 · the brief

Time phone CPR from audio

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

Software that listens to a recorded emergency call and marks the moments the AHA's telecommunicator CPR measures use: when the call is answered, when the call-taker recognizes a cardiac arrest, and when the first chest compression is instructed. It also checks whether the call-taker coached depth, rate and letting the chest rise. Test it on simulated calls in which an actor kneels at a training manikin and a trainee call-taker reads a standard CPR script over a line that reaches only the test room.

Why it is still open

These audits are still done by people listening to recordings. A 2022 study by S. R. Dowker and colleagues built a standard form for that manual review and used it on 65 calls, finding a median of 76 seconds to recognition and 198 seconds to the first instructed compression. Vendors of call-center quality software advertise automatic scoring of calls against protocols, but no published test of such scoring on CPR timing was found.

ebi.ac.ukknogin.com

Built with what exists

speech-to-text with word-level timestamps, such as Whisper with forced alignment; a general AI model that tags each event against the AHA measure definitions, or one built for fixed-answer questions such as Gemini Decision 1 (see the Note "A thousand questions, one pass"); manikins that log each compression, such as Laerdal's QCPR models, to record when compressions really began.

  1. First step

    Record 20 simulated calls with volunteers and a published dispatcher CPR script, and have two reviewers time them using the measure definitions from Dowker and colleagues.

  2. Done looks like

    On 200 simulated calls, the software's times for recognition and first instructed compression fall within 5 seconds of two trained human reviewers in at least 90% of calls, its coaching checks agree with the reviewers on at least 90% of items, and the code and simulated calls are released.