Google says a new Gemini-powered tool is helping Watch Duty turn noisy firefighting radio into searchable text so its volunteer reporters can find possible wildfire dispatches sooner. The project was built with support from a Google.org Fellowship after six months of work alongside Watch Duty, a nonprofit service run by current and former firefighters, dispatchers and first responders. Its purpose is specific: make difficult radio audio easier to review, not publish an emergency alert without a person checking the information.

Why fire radio is difficult to monitor

In many areas, the first reports of a new incident still travel over analog VHF radio. A dispatch may be mixed with static, engine noise, wind and short forms of technical language that are familiar to emergency crews but hard for an outside listener to parse. Google says Watch Duty reporters have monitored as many as 20 feeds manually. That arrangement can leave a small team listening across several channels while trying to separate a new incident from routine traffic.

For the tool, the Google team says it used more than 20,000 hours of earlier radio recordings. The system is described as reducing heavy static and converting specialized firefighting terminology into clean text. That gives reporters a way to search and scan transmissions instead of relying entirely on continuous listening. It also creates a transcript that can be reviewed against what was actually said, rather than treating a model's interpretation as a verified incident report.

The performance figures Google shared

Google reports that the tool lowered word error rate on degraded VHF audio from 38 percent to 22.34 percent. That is a 15.66 percentage-point decrease, or about 41 percent fewer word errors relative to the earlier figure. The company also says the system now transcribes more than 200 hours of radio speech per day across roughly 400 active feeds. Google says it has been designed and tested to handle more than 10,000 feeds, a capacity figure that should not be confused with the number currently running.

A lower transcription error rate can make a noisy message easier to understand, but it does not establish that every name, location or tactical instruction is correct. The headline percentage describes word recognition, not the accuracy of a complete fire alert. Context still matters: radio traffic can contain overlapping voices, local place names and shorthand whose meaning depends on the incident.

A faster signal still goes through reporters

Google points to the July 2026 Railroad Fire in Washington as an example. The tool transcribed a 911 dispatch buried in static and flagged it to Watch Duty in under 30 seconds. Reporters then verified the incident before the organization sent its first nighttime Critical Alert. Google also says reporter satisfaction with radio-transcription tools rose from 35.3 percent to 92.3 percent after the collaboration. These are figures shared by Google, rather than an independent evaluation published with the announcement.

The design reflects a useful boundary for AI in emergency communications. Software can listen continuously and surface a likely signal at machine speed; trained people can confirm whether the transcript points to a real event and whether a public alert is justified. The practical measure is not simply how quickly a model produces text, but whether it helps a small reporting team notice important information sooner without adding misleading alarms.

FILEZUNA CONTEXT

Why it matters

This announcement is strongest as a description of an operational assistive tool, not a claim that AI now dispatches emergency warnings. The reported error-rate improvement is meaningful, yet transcription quality and incident verification are separate steps: a system can recognize most words and still misread the one location or instruction that matters. Google says Watch Duty reporters verify incidents before alerts, and the Railroad Fire example follows that pattern. The next useful evidence would be an independently documented evaluation across different regions, accents, radio conditions and false-alarm rates. Until then, the figures show promising performance in one real workflow while remaining company-reported results.

SOURCE NOTES

Sources and further reading

This story is written in its own words. These links identify the reporting and announcements used to verify the details above.

Facts checked Nov 21, 2026.

FZEND OF STORY
More Technology stories