Industries / Industrial equipment and services / Industrial services / Safety training
Language and literacy
Teaching safety to workers who do not speak the site's main language well, or who read little in any language.
6
Milestones achieved, from 1994 to 2024.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1994
A manual for low literacy
Elizabeth Szudy and Michele Gonzalez Arroyo of the Labor Occupational Health Program at the University of California, Berkeley, published The Right to Understand, a manual for safety trainers whose students read little. It shows how to write materials that are easy to read, check them with readability formulas and focus groups, run participatory classes, and test skills in ways that need little reading or writing.
Credit
Elizabeth Szudy and Michele Gonzalez Arroyo, Labor Occupational Health Program, UC Berkeley
1998
Films without words
After a European Commission film festival in Strasbourg in 1995, four safety communicators from France, Germany, Sweden and the UK set out to make one film usable across Europe, because films made in one country were hard to adapt for others. The Strasbourg studio Via Storia created a cartoon worker, Napo, and the first film, Best Signs Story, was shown at the EU film festival in Edinburgh in 1998. EU-OSHA, the EU's workplace safety agency, calls the series "language-free films"; the consortium behind them was dissolved at the end of 2025, and the existing films stay available.
LessonA film with no words crosses borders, but its makers say it should not "be seen as training or teaching films", only as a way into a topic.
Credit
the Napo Consortium (at its end AUVA, DGUV, EU-OSHA, INAIL and TNO) and Via Storia
2010
Training must be understood
On 28 April 2010 David Michaels, head of OSHA (the US Occupational Safety and Health Administration), told its regional offices that all training its standards require must be given "in a manner that employees can understand": in a language and vocabulary they know, and not as reading matter for workers who cannot read. Inspectors were told to look past training records and check whether workers could understand what they had been taught. The memo cites a 2002 federal appeals court ruling that an employer cannot rely on a work rule it never explained in a language the worker understood.
Credit
OSHA
2010
Spill training in three languages
After the Deepwater Horizon oil spill in the Gulf of Mexico, BP trained more than 100,000 cleanup workers in English, Spanish and Vietnamese under requirements set by OSHA, whose staff fluent in Spanish and Vietnamese watched sessions and interviewed workers, by the account of the heads of OSHA and NIOSH (the US worker safety research institute). A 2018 policy statement of the American Public Health Association, citing a 2013 conference paper by Sandra Sarpy, says Vietnamese workers who were trained in another language rated the training less effective and later reported fewer safe practices.
LessonTranslation is not enough on its own; the statement names poor translation and a lack of familiar examples.
Credit
OSHA, NIOSH, NIEHS and BP
2010
Peer trainers for day laborers
Researchers surveyed more than 300 Latino day laborers in New Jersey before a one-day Spanish-language class of at least six hours, led by trained worker-trainers and built around group problem solving. Two to six months later, the 70 men who could be reached reported more use of hard hats, steel-toe boots, harnesses and safety vests, and 66% had shared the class workbook with friends or co-workers.
LessonOnly 22% of these mobile workers could be found again, so the results rest on few people.
Credit
Quinn Williams Jr., Michele Ochsner and colleagues
2024
Japan's trainees hurt more often
In 2024 Japan counted 3.98 deaths and injuries a year per 1,000 technical intern trainees, foreign workers admitted on a training visa, against 2.3 for all workers; in construction the figures were 6.20 against 4.2. The numbers are the health ministry's, compiled for its 2025 employer seminars. The ministry names short job experience, limited Japanese and poor passing-on of hazards among the causes, and publishes manga (comic-strip) training materials for newcomers in up to 14 languages, for more than 20 industries and tasks.
Credit
Japan's Ministry of Health, Labour and Welfare (MHLW); Tokyo Labour Standards Association Federation
Watch for
Translation app for site terms
Crowdworks AI says Safety Talk, a translation app it built with a large South Korean builder it calls Company S, covers 22 languages, draws on more than 100,000 construction and shipbuilding terms from a government glossary and worker interviews, and reached over 90% translation accuracy in field tests on 1.05 million data samples. The same account says Google Translate and Papago, a Korean app widely used on sites, mistranslated industry terms.
Achieved whenan outside group publishes accuracy on safety-critical terms and sentences, by language, rated by native speakers, with the test set described.
Claimed by
Crowdworks AI, in an ITU report correct as of June 2025
Live translation at every site
Dongbu Construction says it has put an AI simultaneous translation program on all its South Korean construction sites, for pre-work toolbox meetings, risk assessment briefings and safety training for new hires, and that it helps foreign workers understand job hazards accurately.
Achieved whenan outside check, such as a comprehension test of foreign workers after translated briefings, reports results.
Claimed by
Dongbu Construction, July 2026
A phrasebook app for Spanish
Clemson University researchers built Habla Seguro, an app of checked bilingual construction safety phrases with audio and offline use. They report that 31 English-speaking supervisors and 31 Spanish-speaking workers rated it above neutral on accuracy, navigation, cultural fit and audio clarity, and that it gave faster and more precise translations than Google Translate.
Achieved whensomeone other than the developers measures what workers understood from the app's phrases, not only how they rated it.
Claimed by
Adrian Cappella, Shima Clarke, Jason Lucas and Joseph Burgett, Clemson University, 2026
Open
01
Translated safety talks, scored
A public test of machine translation on toolbox talks, the short safety talks given before work. Take 30 of the free construction toolbox talks published by CPWR, the US construction safety research center, which come in English and in Spanish versions made by CPWR. Translate the English into Spanish and four other languages spoken on local sites with three systems, and have two native speakers per language mark every sentence whose safety meaning is lost or changed, such as a dropped "not", a wrong unit or a wrong tool name.
The briefCredit · unclaimed02
Briefings transcribed in site noise
A public set of recorded safety briefings for testing speech-to-text, the software that writes down speech, in the noise of a work site. Native speakers of six languages read scripted toolbox talks; the recordings are mixed with recorded jackhammer, drill and engine noise at set loudness ratios; each system's word error rate (the share of words it gets wrong) and its error rate on safety terms are measured.
The briefCredit · unclaimed
Open 01 · the brief
Translated safety talks, scored
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of machine translation on toolbox talks, the short safety talks given before work. Take 30 of the free construction toolbox talks published by CPWR, the US construction safety research center, which come in English and in Spanish versions made by CPWR. Translate the English into Spanish and four other languages spoken on local sites with three systems, and have two native speakers per language mark every sentence whose safety meaning is lost or changed, such as a dropped "not", a wrong unit or a wrong tool name.
Why it is still open
The nearest measured test is from emergency medicine: Google Translate kept the overall meaning of 82.5% of 400 discharge sentences, from 94% in Spanish down to 55% in Armenian (Taira and colleagues, 2021). For construction, what has been published is the developers' own usability ratings for one Spanish phrase app (Habla Seguro, 2026) and a vendor's own accuracy figure (Crowdworks AI, 2025). A search found no test of machine-translated safety training rated for safety-critical errors.
Built with what exists
CPWR's English and Spanish toolbox talks; DeepL, Google Translate and an open model such as Mistral Large 4 (see the Note "A trillion open parameters"); the error types of Multidimensional Quality Metrics (MQM), a shared list of translation errors and their severity; a shared spreadsheet for the ratings.
First step
Translate five CPWR talks into Spanish with the three systems and have two bilingual raters compare them line by line with CPWR's Spanish versions.
Done looks like
30 talks in five languages from three systems, every sentence rated by two native speakers who agree on at least 80% of their marks; a published count of safety-critical errors per 100 sentences for each system and language, with CPWR's own Spanish scored the same way as a reference.
Open 02 · the brief
Briefings transcribed in site noise
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of recorded safety briefings for testing speech-to-text, the software that writes down speech, in the noise of a work site. Native speakers of six languages read scripted toolbox talks; the recordings are mixed with recorded jackhammer, drill and engine noise at set loudness ratios; each system's word error rate (the share of words it gets wrong) and its error rate on safety terms are measured.
Why it is still open
Live translation on sites starts with speech-to-text, and both the Crowdworks AI and Habla Seguro accounts name noisy sites as a weakness still to be fixed. The nearest public set, RescueSpeech (2023), holds German speech from simulated rescue exercises, and its authors found the best systems "still far from reaching an acceptable level". Denis Moser, Nikola Stanic and Murat Sariyar of Bern University of Applied Sciences built this kind of test for German emergency medical dialogues (2025): 99 synthetic dialogues mixed with crowd, traffic and ambulance noise at five levels, scored with an error rate weighted toward medical terms. The SAFE-T corpus of the Linguistic Data Consortium (2020) holds over 300 hours of English from 115 speakers doing simulated public safety tasks with background noise. A search found no set of safety briefings recorded with construction or plant noise, or in the languages spoken on those sites.
Built with what exists
ElevenLabs Scribe v3 (see the Note "Fewer word errors") and OpenAI's open Whisper large-v3 model; noise clips from UrbanSound8K, a public sound set with jackhammer and drilling classes; CPWR toolbox talks as scripts; the Python package jiwer to compute error rates.
First step
Record four Spanish speakers reading three CPWR talks, mix in drilling and jackhammer clips at two levels, and score two systems.
Done looks like
At least 2 hours of speech per language from at least four speakers each, mixed at three noise levels; published word and safety-term error rates for at least three systems; audio and transcripts released under an open license.