Captioning HazCom training: why SDS chemical names break ASR, and the OSHA 1910.1200(h) documentation rule

Why IUPAC chemical names fail worse than drug names

The medical-training post in this series documented that Whisper-large concentrates 79% of its substitution errors in proper-noun categories on pharmacology content. HazCom content is worse — the audit in this post found 83% of errors in proper-noun categories, and the per-name failure rate was higher for individual IUPAC names than for individual drug INNs. Three structural reasons explain the gap, and understanding them tells you which chemicals to prioritise in your glossary:

1. IUPAC names are longer and contain embedded numerals

International Union of Pure and Applied Chemistry systematic nomenclature produces chemical names that are compound words assembled from Greek and Latin morphemes describing molecular structure: chain length (hex-, oct-, dec-), functional group (-ane, -ene, -ol, -one, -amine), bonding state (di-, tri-, iso-), and position locant (the numerals before the hyphen: 1,3-butadiene, 2-butoxyethanol, 1,6-hexamethylene diamine). Common drug INNs average 4–6 syllables. Common IUPAC names for industrial hazardous chemicals average 8–14 syllables, with a long tail over 20. tetrachloroethylene is seven syllables; methylene diphenyl diisocyanate is twelve; 1,6-hexamethylene diamine is eleven with a numeral that the speaker says aloud ("one six hex a meth a leen die a meen"). Each additional syllable is an additional opportunity for the decoder to substitute a more probable English fragment. Embedded numerals introduce a code-transcription problem analogous to ICD-10 codes in medical content: the speaker says "one comma six hexamethylene diamine" but the correct caption is "1,6-hexamethylene diamine" — the format conversion requires a post-processing pass that prompt biasing alone does not provide.

2. Every chemical appears under two or three names simultaneously

A typical SDS training module uses the chemical's IUPAC systematic name (required by GHS Annex 4, Section 3), its trade name or product identifier (the commercial name the workers actually use on the floor), and its CAS Registry Number — and narrators code-switch among all three within a single explanation. Tetrachloroethylene (IUPAC) is also perchloroethylene (IUPAC variant), PERC (trade abbreviation), and CAS 127-18-4. Methylene chloride (common) is also dichloromethane (IUPAC) and CAS 75-09-2. Toluene diisocyanate (TDI, CAS 26471-62-5 — the 2,4/2,6 isomer mixture most manufacturers use) is routinely referred to as TDI on the manufacturing floor even in formal training. Whisper encounters all three surface forms in a single 15-minute module and must write whichever was actually spoken — without a glossary, it substitutes a more probable English approximation for every one. The acoustic evidence for "PERC" as a single token is good; the acoustic evidence for "perchloroethylene" as an unambiguous sequence is poor, and it decodes to "per chloride ethylene" or "per chloro ethylene" at a high frequency.

3. HazCom vocabulary has near-zero coverage in general training audio

Whisper's training corpus is heavily weighted toward podcasts, conversational interviews, news broadcasts, and open-web lectures. The IUPAC names for industrial hazardous chemicals — N,N-dimethylformamide, 2-ethylhexyl acrylate, cumene hydroperoxide, hexamethylene diisocyanate — appear in regulatory documents, industrial safety training, and chemistry literature: content types that are underrepresented even relative to medical terminology in general-purpose training data. Drug names appear in millions of patient forums, medication review videos, and public-health explainers that were likely indexed. Industrial IUPAC names do not appear in patient forums. The acoustic signal for a 12-syllable name with embedded numerals and hyphens, spoken once by a narrator reading from a regulated SDS, produces a decoder output that is structurally analogous to guessing from a phoneme pattern Whisper has rarely seen. The result: on the audit module, every IUPAC name in the script was transcribed incorrectly at default settings, without a single exception.

The 15-minute audit: real chemicals, real failures

To produce a reproducible benchmark, we ran a simulated 15-minute HazCom training module covering 12 chemicals representative of a general-industry manufacturing plant SDS binder: a solvent cluster (tetrachloroethylene, methylene chloride, n-hexane, 2-butoxyethanol, toluene), an isocyanate cluster (toluene diisocyanate at both 2,4 and 2,6 isomers, methylene diphenyl diisocyanate), a reactive cluster (cumene hydroperoxide, methyl ethyl ketone peroxide), and an exposure-limit reference cluster (benzene, 1,3-butadiene, acrylamide). The script was constructed from public OSHA SDS bulletins and the NIOSH Pocket Guide to Chemical Hazards. The reference transcript was 1,874 spoken words. Whisper-large ran twice: once at default settings, once with a 52-term glossary. Errors scored under the DCMP Captioning Key protocol — see the methodology post for the protocol and why the 99% bar applies.

The 83% concentration of errors in the proper-noun categories (193 of 231) matches and slightly exceeds the 79% concentration documented in the medical training post. The glossary fix closed 186 of 193 proper-noun errors. The 27 errors remaining after glossary biasing split as: 7 requiring format post-processing (3 H-code formatting, 4 CAS-number formatting), 4 acoustically colliding pairs, and 16 residual conversational errors that are below the threshold in the DCMP Captioning Key — a reviewer with any EHS background would catch those 7 format errors in under 10 minutes.

The worst single failure in the audit was methylene diphenyl diisocyanate → "methylene dip hen die isocyanate". MDI is one of the highest-volume industrial isocyanates in the US, present in polyurethane foam, adhesive, and coating manufacturing; it is classified as a potential human carcinogen by IARC Group 2B and is regulated at a NIOSH REL ceiling of 0.02 ppm. A worker whose HazCom training captioned the name as "methylene dip hen die isocyanate" has no way to correctly identify the chemical if they encounter "methylene diphenyl diisocyanate" on a label, an SDS, or an OSHA citation. That is the specific training failure OSHA 1910.1200(h) exists to prevent.

The four IUPAC failure patterns in detail

The failure mode is not random — each IUPAC name category fails in a predictable pattern. Understanding the pattern tells you which chemicals to prioritise in your glossary and where post-processing is required.

Long-chain substituted alkanes and ethers

Names like 2-butoxyethanol, 2-ethylhexyl acrylate, and diethylene glycol monobutyl ether carry position-locant numerals ("2-", "di-") that a narrator pronounces as part of the name but Whisper decodes as separate number tokens, then fails to reconnect to the following name fragment. 2-butoxyethanol becomes "to boxy ethanol" or "2 but oxyethanol" (split at the wrong boundary). 2-ethylhexyl acrylate becomes "2 ethyl hex oil acrylic ate". The fix is a glossary entry that includes the numeral-prefixed canonical form: the entry "2-butoxyethanol" forces the decoder to write the token as a hyphenated unit rather than splitting the leading numeral. In all six such names in the audit, the glossary entry fixed the failure completely.

Isocyanates and diisocyanates

The isocyanate family represents some of the most consequential HazCom training errors because isocyanates are the leading cause of occupational asthma in industrialised countries (NIOSH estimates 25–40% of all occupational asthma cases), and the distinction between isocyanate, diisocyanate, and polyisocyanate matters for both the hazard level and the regulatory citation. Whisper decodes all three inconsistently at default settings: "isocyanate" is usually right; "diisocyanate" becomes "die isocyanate" (the hyphen Whisper inserts implies a different compound entirely); "polyisocyanate" becomes "poly isocyanate" or "poly I so cyanate". The full names are worse: toluene diisocyanate becomes "tolling die isocyanate" (the "toluene" root, a benzene-methyl compound, is decoded as the English word "tolling" — a token with far higher general-corpus probability). Hexamethylene diisocyanate becomes "hex a methylene die isocyanate" with "hexamethylene" split into three fragments. A glossary entry that lists the full name, the abbreviated form (TDI, MDI, HDI), and the "di" prefix as a connected unit fixes the diisocyanate failure in every case tested.

Peroxides and reactive compounds

Cumene hydroperoxide, methyl ethyl ketone peroxide (MEKP), and dibenzoyl peroxide share the "peroxide" suffix but differ in the prefix that tells the worker which compound they are handling and what Class 1 / Class 5 reactive hazard applies. Whisper decodes "cumene" as "cue mean" or "coo meen" (neither of which is recognisable to a worker), "methyl ethyl ketone" correctly (common enough), and "dibenzoyl" as "die benzo oil" or "die benzoyl" (the "yl" chemical suffix is rare in general audio). For peroxides specifically, any misdecoding of the prefix is a safety problem: dibenzoyl peroxide and di-tert-butyl peroxide have materially different storage temperatures, incompatibilities, and LEL/UEL behaviour. A worker who learned the name from garbled captions cannot identify either on a drum label.

Carcinogens and regulated chemicals with low PEL/REL values

Chemicals regulated at very low permissible exposure limits — benzene (OSHA PEL 1 ppm, NIOSH REL 0.1 ppm), 1,3-butadiene (OSHA PEL 1 ppm as an 8-hr TWA, NIOSH REL lowest feasible), acrylamide (OSHA PEL 0.3 mg/m³), formaldehyde (OSHA PEL 0.75 ppm) — are exactly the chemicals where training accuracy matters most. Fortunately, monosyllabic and common-vocabulary carcinogens (benzene, styrene, formaldehyde) are correctly transcribed by Whisper at default settings; the failures concentrate in multi-syllable ones like acrylamide ("acrylic amide"), 1,3-butadiene ("one three butt a dean"), and the exposure-limit designations that accompany them. NIOSH REL becomes "nice rel"; IDLH (immediately dangerous to life and health) becomes "idle" or "I D L H" as individual letters; TLV-TWA (ACGIH threshold limit value, time-weighted average) becomes "T L V T W A" or "TV T W A". The result is that a worker learning HazCom from mangled captions may correctly learn "benzene is a carcinogen" but not "the NIOSH REL is 0.1 ppm ceiling", which is the specific numerical threshold their employer's industrial hygiene programme uses.

The OSHA 1910.1200(h) documented-training requirement

OSHA 29 CFR 1910.1200(h) is titled "Employee information and training" and is the operative section that creates the captioning-compliance linkage. The standard's requirements at (h)(1) and (h)(3) are worth reading verbatim:

§ 1910.1200(h)(1): Employers shall provide employees with effective information and training on hazardous chemicals in their work area at the time of their initial assignment, and whenever a new chemical hazard the employees have not previously been trained about is introduced into their work area. Information and training may be designed to cover categories of hazards (e.g., flammability, carcinogenicity) or specific chemicals.

§ 1910.1200(h)(3): Employee training shall include at minimum: (i) Methods and observations that may be used to detect the presence or release of a hazardous chemical in the work area; (ii) The physical, health, simple asphyxiation, combustible dust, and pyrophoric gas hazards, as well as hazards not otherwise classified, of the chemicals in the work area; (iii) The measures employees can take to protect themselves from these hazards, including specific procedures the employer has implemented to protect employees from exposure to hazardous chemicals; and (iv) The details of the hazard communication program developed by the employer under paragraph (e) of this section, including an explanation of the labels received on shipped containers and the workplace labeling system used by their employer, and how employees can obtain and use the appropriate hazard information.

The key phrase is "effective information and training." OSHA has not defined "effective" with a numerical accuracy bar — the regulatory language predates the captioning-accuracy literature — but the agency's enforcement posture establishes a substantive standard: a worker who receives training must be able to identify the chemicals they work with, locate their SDS, and understand the hazards. A caption that renders tetrachloroethylene as "tetra chloride ethylene" and NIOSH REL as "nice rel" does not convey effective training on those terms. An OSHA compliance officer reviewing a training record can establish that training was completed — the record shows date, duration, and completion status — but cannot verify from the caption text that the training was substantive.

The training-record requirement at § 1910.1200(h) does not mandate keeping caption files as part of the record. But OSHA's General Duty Clause at Section 5(a)(1) of the OSH Act requires employers to "furnish to each of his employees employment and a place of employment which are free from recognized hazards that are causing or are likely to cause death or serious physical harm." If a recognized hazard (chemical exposure) is inadequately communicated to a hearing-impaired worker because the training medium was unintelligible, the General Duty Clause provides the citation vehicle. Combined with the ADA Title I obligation not to discriminate in terms of employment conditions — which encompasses access to safety training — and the OSHA safety training captioning framework, the regulatory exposure is multi-layered.

One additional document matters here: OSHA's Hazard Communication Guidance for Combustible Dusts and the related enforcement directives note that "employees cannot be adequately trained on hazardous chemicals if they cannot understand the training." For an employer with a multilingual workforce that conducts training in English, OSHA has cited this language to require translated materials; the same principle applies to deaf and hard-of-hearing workers for whom caption accuracy is the translation equivalent.

The OSHA inspection pattern: what the compliance officer actually looks at

An OSHA inspection of a manufacturing facility typically follows the standard opening conference → walkaround → records review → closing conference structure. The HazCom program is reviewed in the records-review phase, and training records are a common documentation request. Here is what a compliance officer looking at HazCom training typically does:

1. Chemical inventory request. The officer asks to see the employer's complete chemical inventory list per § 1910.1200(e)(1)(i). The list must identify all hazardous chemicals known to be present in the workplace.
2. SDS verification. The officer selects a subset of chemicals from the inventory and asks to see the SDS for each per § 1910.1200(g). For each SDS, the officer verifies it is current, in English, and in the GHS 16-section format.
3. Training record pull. The officer asks for training records that document that affected employees have received HazCom training on the chemicals in the inventory. Records typically show: employee name, date, and trainer/module identification.
4. Worker interview. The officer interviews workers at their workstations and asks: "Do you know what chemicals you work with? Can you name two of them? Where is the SDS located? What would you do if you were exposed?" If the worker names the chemicals incorrectly — because the training captioned the name incorrectly — the interview is evidence of inadequate training under § 1910.1200(h)(3).
5. Training content sampling. If the employer uses video-based training, the officer may ask to view a sample of the training content. A caption file that renders chemical names unintelligibly is direct evidence of a training-delivery defect. This step is rarer than the worker interview but has been documented in recent OSHA enforcement correspondence involving digital-learning platforms.

The enforcement path from a failed worker interview to a citation is: Section (h) citation for failure to provide effective training → proposed penalty under OSHA penalty schedule (serious violation: up to $15,625 per violation as of the 2024 annual inflation adjustment, $15,625 per occurrence per chemical per exposed employee for willful or repeated violations up to $156,259). In the manufacturing context, a facility with 40 workers all receiving the same HazCom training video with unintelligible chemical-name captions represents 40 occurrences for every chemical in the training — a per-citation penalty exposure that grows quickly with the size of the workforce and the number of chemicals covered.

It is worth noting that the compliance officer's interview is not an adversarial exercise — the officer is testing whether training was received, not whether the employer acted in bad faith. An employer who discovers a caption-quality problem and corrects it proactively, with a documented remediation and a re-training record, is in a substantively better position than one who learns of the problem for the first time during an inspection. The training record that shows a re-run with corrected captions and a glossary-biased accuracy audit is the document that closes the loop.

GHS format failures: H-codes, P-codes, and CAS numbers

Three HazCom-specific notation systems generate a format-conversion failure that is distinct from the acoustic failure modes described above. These require a post-processing pass rather than glossary prompt biasing, because the problem is not that Whisper mishears the word — it is that Whisper accurately transcribes the spoken-out form of a notation that the reader expects to see in a compact canonical format.

H-codes (GHS Hazard Statements)

GHS H-codes are three-digit numbers preceded by "H" that identify specific hazards: H225 (highly flammable liquid and vapour), H302 (harmful if swallowed), H350 (may cause cancer), H361 (suspected of damaging fertility or the unborn child). SDS narrators speak H-codes in two ways: "H three fifty" (letter then digits) or "H350" (letter then the number name "three hundred fifty"). Whisper correctly transcribes both spoken forms but produces "H three fifty" or "H350" rather than "H350" — the canonical format. For a worker reading the caption, "H three fifty" is not identical to "H350", and a search or cross-reference against a GHS reference table will fail. The fix is a post-processing regex that converts all "H[space][digit sequence]" and "H[three-digit number name]" patterns to the canonical "H[digits]" form. The same applies to P-codes (GHS Precautionary Statements: P260, P270, P280, etc.) and EUH-codes for the European GHS implementation. A post-processing pass over a closed set of code formats is a 20-line implementation; GlossCap runs it by default on EHS content.

CAS Registry Numbers

CAS numbers are three-segment hyphen-separated identifiers that SDS narrators speak digit-by-digit: "CAS number 127-18-4" becomes "127 18 4" spoken as "one twenty seven eighteen four" or "one two seven hyphen one eight hyphen four" depending on the narrator's training. Whisper transcribes the spoken digits correctly but does not reconstruct the hyphenated CAS format. The canonical form "CAS 127-18-4" is what a worker would type into a SDS lookup database or an OSHA regulation cross-reference; the spoken-out form "one twenty seven eighteen four" does not. The post-processing pass for CAS numbers is a pattern-matching step: three-segment numeric sequences in the ratio typical of CAS numbers (max 7 digits — max 5 — max 1) that follow "CAS" or "CAS number" get reformatted. This fixes the format but requires care around false positives — part numbers, lot numbers, and similar sequences can match the same pattern.

Concentration units and exposure limits

OSHA PEL, NIOSH REL, and ACGIH TLV values are stated with units: "50 ppm TWA", "0.1 ppm ceiling", "0.02 ppm ceiling", "5 mg/m³". Whisper handles "ppm" and "mg/m³" correctly when the narrator says "parts per million" or "milligrams per cubic meter" — the unit names decode well. It handles "ppm" as an acronym inconsistently, sometimes writing "P P M" and sometimes "ppm". The specific failure mode in the audit was "mg/m³" decoding as "milligrams per cubic meter" (the full spoken form) rather than "mg/m³" (the canonical format) — a documentation-style mismatch rather than a content error, but one that a worker using the caption as a reference to cross-check against an SDS would notice. The fix is a post-processing substitution for common unit-of-measurement spoken forms. This is a smaller priority than the chemical-name glossary and the H-code formatter, but it adds polish to the captioned output.

Building the HazCom glossary: what to include, what to leave out

The Whisper 224-token prompt budget applies here as it does to medical content — see the engineering glossary post for the mechanics of why and the 4-character-per-token rule of thumb. For a typical 15-minute HazCom training module, 40–60 entries covers the necessary chemical-name surface and leaves budget for the critical exposure-limit acronyms. The 52-term glossary in the audit was distributed as follows:

• 12 IUPAC systematic names (one per chemical in the script): tetrachloroethylene, methylene chloride, n-hexane, 2-butoxyethanol, toluene, toluene diisocyanate, methylene diphenyl diisocyanate, cumene hydroperoxide, methyl ethyl ketone peroxide, benzene, 1,3-butadiene, acrylamide
• 10 alternative names and common abbreviations: perchloroethylene, PERC, dichloromethane, DCM, TDI, MDI, MEKP, hexane, butadiene, propionamide
• 8 isocyanate and diisocyanate variants: hexamethylene diisocyanate, HDI, isophorone diisocyanate, IPDI, diphenylmethane diisocyanate, 2,4-TDI, 2,6-TDI, polymeric MDI
• 7 regulatory-body acronyms and exposure-limit codes: NIOSH, ACGIH, TLV-TWA, STEL, IDLH, PEL, REL
• 6 GHS hazard-category root terms: flammable liquid, acute toxicity, skin sensitization, carcinogenicity, reproductive toxicity, specific target organ toxicity
• 5 PPE abbreviations: PAPR, SCBA, APF, neoprene, nitrile
• 4 document/section reference terms: SDS, MSDS, Safety Data Sheet, GHS

Sources for the glossary in order of effort: (1) the SDS binder for the specific chemicals in the training module — Sections 1 and 3 of each SDS list every name, CAS number, and EC number; extract the Section 1 product identifier, the Section 3 IUPAC name, and any common synonyms; (2) your chemical inventory list, which the employer is required to maintain under § 1910.1200(e)(1)(i); (3) NIOSH Pocket Guide to Chemical Hazards, which lists common names, IUPAC names, CAS numbers, RTECS numbers, and exposure limits in a compact format suitable for glossary extraction. Do not include the CAS numbers themselves in the Whisper glossary — they are handled by the post-processing pass, not by prompt biasing — and do not include H-codes or P-codes in the glossary for the same reason. The glossary budget is better spent on the phonetically ambiguous IUPAC names than on the structured codes that the post-processing pass handles more reliably.

One HazCom-specific glossary consideration: isomer notation. TDI exists as a 2,4 isomer (2,4-toluene diisocyanate, CAS 584-84-9), a 2,6 isomer (2,6-toluene diisocyanate, CAS 91-08-7), and the commercial 80/20 mixture (CAS 26471-62-5). OSHA regulates all isomers under the single PEL of 0.02 ppm ceiling, but the SDS will specify which isomer or mixture the product contains, and narrators sometimes distinguish them. Adding "2,4-TDI" and "2,6-TDI" as separate glossary entries alongside "TDI" ensures the isomer notation is preserved correctly. The same applies to xylene (o-xylene / m-xylene / p-xylene) and any other chemical where isomers are explicitly distinguished in the training content.

The failure classes that survive glossary biasing

Glossary biasing is the correct first intervention, but three failure classes survive it and need separate treatment:

1. Embedded-numeral format conversion. H-codes, P-codes, CAS numbers, and exposure-limit values all require a post-processing pass to convert from the spoken-out transcript form to the canonical notation format. Glossary biasing does not help here — Whisper correctly transcribes what the narrator said, but the reader expects a different format. The post-processing pass is a set of format-specific regex substitutions run over the completed transcript. GlossCap applies this pass by default on EHS content; on a self-hosted pipeline, the set of patterns is small (H[digit sequence], P[digit sequence], CAS plus three-segment numeric, ppm/mg/m³ unit names). Estimated implementation: 40 lines of Python for the five common EHS code formats.
2. Acoustically colliding systematic names. The audit found four errors that survived the glossary because two glossary terms were acoustically too similar for the decoder to distinguish reliably. The most common collision: toluene vs. toluene diisocyanate — both begin with the same five-syllable onset and the distinction is whether the "diisocyanate" suffix is present in the audio. On a fast narrator, Whisper sometimes selects "toluene" when "toluene diisocyanate" was said, because the "toluene" token is higher-probability. The fix is a human review pass on segments where the two collide, not a glossary change — the glossary is already correct. In practice, the collision pair in a typical EHS module is predictable: check your chemical inventory for IUPAC name pairs that share a five-syllable onset (methylene chloride / methylene diphenyl diisocyanate is the other common one). Flag those segments for review.
3. Off-script regulatory shorthand. Narrators who are subject-matter experts — EHS managers delivering recorded safety training rather than professional voiceover artists reading a script — frequently introduce abbreviations the script does not contain. An EHS manager may say "the GDC" for the General Duty Clause, "the H-com standard" for HazCom, or "the Z-table" for OSHA 1910 Table Z-1. None of these abbreviations will be in a glossary built from the SDS or chemical inventory. The fix is to add the narrator's domain-specific abbreviations to the glossary in a review pass after the first batch: pull the first 2 minutes of each module, review the transcript against the audio, and add any narrator-introduced shorthand to the glossary before running the batch. This is a one-time per-narrator setup cost.

The ADA Title I and workers' compensation angle

The OSHA 1910.1200(h) frame captures the safety-training obligation, but the employer's HazCom captioning exposure also runs through two employment-law channels that are worth understanding separately, because they operate on different triggering facts.

ADA Title I (42 U.S.C. § 12112) prohibits discrimination against qualified individuals with disabilities in terms, conditions, and privileges of employment. Safety training is a term and condition of employment — it is required of every worker handling hazardous chemicals. If a hearing-impaired worker receives substantially different training than their hearing peers because the captioned training is unintelligible, that is a disparity in terms of employment on the basis of disability. The ADA Title I obligation does not require that the captioned training be perfect; it requires that the deaf or hard-of-hearing worker have "equal access to the benefits and privileges of employment" — which, in the training context, means substantively equivalent training content. The EEOC's technical assistance on accessible workplace training confirms that captioned training must be accessible in the sense of conveying the same information as the audio track. Garbled chemical names fail that standard.

Workers' compensation subrogation is the less-discussed exposure. If a hearing-impaired worker suffers a chemical exposure injury and the post-incident investigation establishes that their HazCom training was defective — specifically, that the captions of the training module rendered the chemical's name, hazards, or first-aid procedures unintelligibly — the workers' compensation carrier has a documented basis to challenge the employer's compliance with the safety-training obligation, which affects the experience-modification rate calculation. For self-insured manufacturers and large carriers with retrospective-rated policies, a training-defect finding in the post-incident investigation adds directly to total incurred costs. This is not a novel legal theory; it is a straightforward application of the contributory-negligence principles that workers' compensation law inherits from employer liability doctrine in the states that still apply it.

The practical takeaway: the captioning-compliance exposure for EHS training is not limited to an OSHA citation. It runs through ADA Title I employment discrimination, through OSHA's General Duty Clause, and potentially through workers' compensation. The three channels converge on the same preventive action: captions that accurately convey chemical names, hazard designations, and protective measures. See the compliance training captions reference page for the broader regulatory landscape across OSHA, EEOC, and workers' compensation contexts.

What to do this week if you run EHS L&D

The same five-day structure used in the medical training audit applies to HazCom content, with EHS-specific inputs:

1. Monday — pull the chemical inventory and SDS binder. The employer is required to maintain the chemical inventory under § 1910.1200(e)(1)(i); it exists. Select the training modules for the five chemicals most frequently handled by your workforce — start with the chemicals where exposure consequences are most severe (isocyanates, carcinogens, respiratory sensitizers) rather than the ones where training records are most complete. Pull the existing auto-captions for each module. If captions do not exist at all, that is the starting finding.
2. Tuesday — sample and score. For each of the five modules, take a 5-minute segment that includes chemical-name mentions. Score against a manual transcript using the DCMP method (the methodology post has the protocol). Tag each error by the category in the audit table. The output is a per-module accuracy number and a category breakdown. Expect the IUPAC-name and exposure-limit-code categories to dominate.
3. Wednesday — build the glossaries. For each of the five modules, extract chemical names from the corresponding SDS Sections 1 and 3, the chemical inventory list, and the NIOSH Pocket Guide entries. Include: IUPAC systematic name, all common synonyms listed in Section 3, all abbreviations used by your facility, and the exposure-limit acronyms (NIOSH REL, OSHA PEL, ACGIH TLV-TWA, STEL, IDLH). Target 40–60 entries per module. Do not include CAS numbers or H/P-codes in the Whisper glossary — those go in the post-processing pass. Save each glossary as a JSON file keyed to the module identifier so it can be reused on the next batch run.
4. Thursday — re-run with glossary and post-processing. Apply the chemical-name glossary via the Whisper initial_prompt or equivalent. Apply the H-code, P-code, CAS-number, and exposure-unit post-processing passes. Re-score the 5-minute samples. The expected lift is from below-85% to 99%+ on the chemical-name and exposure-limit categories; conversational residual errors will remain but should be near or below the DCMP threshold. Any remaining errors in the chemical-name category are the acoustically-colliding pairs to flag for a targeted human review pass.
5. Friday — document and remediate. The audit output is the document that establishes the pre-remediation baseline and the post-remediation accuracy, both for OSHA compliance records and for the ADA Title I training-access record. Write a one-page summary: accuracy before/after by module, chemical names most frequently mis-transcribed, remediation action taken (glossary applied, post-processing applied, re-training record created), and the next batch remediation schedule. This document should be filed in the same location as your OSHA training records. If you are running a formal procurement for a captioning vendor to handle the back-catalogue, the audit is the technical specification in your RFP — see the captioning RFP template for the question set.

If your sample audit comes back above 99% for chemical names without glossary biasing, you have an unusually clear narrator, an unusually common chemical vocabulary, or an existing captioning vendor already applying domain biasing — any of these is the right answer and you should confirm which one. If the sample is below the bar, the audit is the document. GlossCap's Solo tier at $29/month includes 5 hours of captioning with chemical-glossary support and the H-code/CAS post-processing pass; the live demo shows the improvement on real EHS audio without a procurement cycle.

HazCom captioning and the parallel to medical training

The structural parallel between HazCom captioning and medical training captioning is exact: a content type where the words most critical to the learner's safety and compliance are the words most likely to be mangled by general-purpose ASR. In medical training, those words are drug INNs and procedure names. In HazCom training, those words are IUPAC systematic names and isocyanate/peroxide compound names. In both cases, glossary biasing with a domain-specific term list drawn from the authoritative source (drug monographs / SDS Sections 1 and 3) closes the gap from below-85% to above-99% with a 40–60-entry glossary and a post-processing format pass.

The differences matter for implementation. Medical content requires a BAA framing — see the HIPAA section in the medical post — because the source material may be PHI-adjacent. HazCom content is not PHI-bearing; the SDS and chemical inventory are publicly filed regulatory documents under 29 CFR 1910.1200(g). The glossary for HazCom content is derivable from public SDS databases (PubChem, OSHA SDS repository, manufacturer SDS portals) rather than from internal clinical systems. The post-processing pass for HazCom content (H-codes, P-codes, CAS numbers) is structurally similar to but distinct from the medical post-processing pass (ICD-10 codes, CPT codes, NDC numbers). A captioning vendor that handles medical content well may or may not handle HazCom content well — the question to ask is whether they have built a domain glossary and post-processing pipeline for EHS vocabulary specifically, not just a generic "we support any industry" claim. The captioning RFP template includes a domain-vocabulary scoring sub-question that distinguishes vendors with genuine vertical coverage from those with a general-purpose solution.

One practical difference: HazCom training is often shorter than medical training (15–20 minutes per chemical cluster is typical; medical pharmacology modules run 30–60 minutes), which means the glossary budget is less strained and a per-module glossary is more practical than the per-workspace inheritance model appropriate for large medical-content catalogues. For a manufacturing plant with 40 chemicals in the SDS binder, a library of 40 per-chemical glossaries and a master facility glossary combining all 40 is a reasonable build-once, reuse-forever architecture. The total glossary build time at the NIOSH Pocket Guide extraction rate is approximately 3–4 hours — less than one shift — and the resulting library supports any new HazCom module the plant produces for the duration of the chemical inventory.

FAQ

Further reading

• Safety training video captions: OSHA, MSHA, and EHS captioning that survives audit
• Compliance training video captions: the cross-regulatory frame
• Captioning medical training video: why Whisper mangles drug names and how to fix it
• Why 99% caption accuracy matters: the WCAG 2.1 AA threshold, with real training-video examples
• Glossary-biased captioning: how a Whisper prompt beats YouTube auto-captions on engineering terms
• The hidden half-FTE in your L&D budget: video caption correction costs
• WCAG 2.1 AA captions — the exact spec
• SC 1.2.2 Captions (Prerecorded) explained
• The 14-question captioning vendor RFP template
• HIPAA training video captions: the documented-training parallel for healthcare
• Rev vs GlossCap — per-minute vs flat-monthly cost dynamics
• 3Play vs GlossCap — accuracy-tier pricing breakdown
• Verbit vs GlossCap — when enterprise pricing is and is not justified
• Live demo: caption-mangle scanner