Google's new AI transcribes 85 languages at near-professional accuracy. That's the toolbox-talk gap for a third of the construction workforce

Google shipped a new speech-to-text model on August 26 that reports a 2.6% word error rate across more than 85 languages, including Spanish. Nearly a third of the U.S. construction workforce is Hispanic. Put those two facts next to each other and you get a real documentation gap: most jobsite safety talks, daily logs, and incident reports still get filtered through translation, informal paraphrasing, or don't get written down in a worker's own words at all.

What did Google actually release?

Gemini 3.5 Transcribe is DeepMind's newest transcription model, built to handle real, messy speech rather than clean studio audio. Google reports a 2.6% word error rate on recorded audio and 4.0% on live streaming, roughly 70% lower latency than its predecessor Chirp 3, and automatic language detection across 85-plus languages. It strips filler words, corrects mid-sentence corrections, and formats the output on its own. It ships as two interfaces: a live API for real-time streaming and an interactions API for recorded audio with speaker attribution and timestamps — the second one is what matters for a documented safety record.

Why is this a different story than the last few voice-AI pieces?

The last three AI voice tools covered here — OpenAI's Ultrafast tier, xAI's Grok Voice, and OpenAI's full-duplex GPT Live — all solved the same problem: latency. Getting a hands-free daily log dictated in under a second so a super doesn't have to stop and type. That's a real workflow win, but it assumes the person talking is dictating in English.

Gemini 3.5 Transcribe's number isn't about speed. It's about whether the model can accurately turn speech in a language other than English into usable text at all. That's a different, and on current workforce numbers, larger problem.

How big is the language gap on an actual jobsite?

Toolbox talks, daily field reports, and incident interviews are exactly the kind of speech this model targets — spoken, informal, often given by someone whose first language isn't English, to a system that today either isn't capturing it at all or is capturing it through a bilingual supervisor's paraphrase.

What would actually change for a GC or sub?

Three documentation gaps this specifically targets:

1. Toolbox talks. A Spanish-speaking foreman delivers the talk in Spanish; the transcript exists in his own words, timestamped, instead of relying on a translated script nobody confirms was followed.
2. Daily field reports. A crew lead who isn't comfortable writing in English can dictate directly instead of relaying through a bilingual peer or skipping the log entirely.
3. Incident interviews. A worker's own account gets captured in their language before it passes through translation, preserving the original statement for the record.

What shouldn't change yet?

Google's accuracy numbers are Google's own benchmark, not an independent test on jobsite audio — generator noise, radios, wind, and regional accents are a harder environment than a clean benchmark set. And a transcript isn't a certified translation or a legal document. Anything headed into an official safety file, an incident report, or an OSHA response still needs a bilingual supervisor to check it before it's relied on. Pilot this on one non-critical use — a weekly toolbox talk log, not an incident report — and compare the output against a known transcript before trusting it with anything that has legal weight.

This is a different problem than the one solved in OpenAI's ultrafast voice tier for hands-free field logging — that piece was about speed for English dictation; this one is about accuracy for everyone else on the crew.

Forward this to whoever runs your firm's safety documentation.

Construction AI Brief publishes a new piece on AI's construction angle most days. Subscribe at constructionaibrief.com.