AI NewsModels & agentsAnnouncement

Watch Duty uses Gemini to catch wildfire dispatches in under 30 seconds

Google built a Gemini tool with the wildfire alert nonprofit Watch Duty. It flags a new fire dispatch in under 30 seconds.

AI News

Editorial2 min read

LinkedInX
Blog illustration for a story about AI reading fire radio

Why it mattersA production transcription job can win on a narrow audio domain when the vendor commits engineers to it. Google quotes measured accuracy, not a demo.

The hardest part of a wildfire alert is not the model. It is the fire radio. The sound is noisy, voices overlap, and local code names are not in any general speech model. Google says a team of its engineers worked full time with Watch Duty for six months. Together they built a Gemini tool that reads those radio feeds live.

Watch Duty is a nonprofit alert app. It is run by active and retired firefighters, dispatchers, and first responders. Its volunteers listen to fire radio and write push alerts. Before the tool, a volunteer had to hear each dispatch and decide if it was a new fire. Then they typed out the location and sent the alert by hand. Google says the new tool now sits between the radio and those volunteers. It turns the speech into clean, searchable text.

The numbers Google gives

Google says the tool reduced the word error rate on degraded VHF audio from 38 percent to 22.34 percent. It now transcribes more than 200 hours of radio speech a day. It covers about 400 active feeds. The team trained it on more than 20,000 hours of past radio recordings. Google says the system flagged the Railroad Fire dispatch in under 30 seconds. Google says Watch Duty reached more than 16 million people last year. No independent benchmark is given.

The model is described as Gemini. No version is named. The deployment runs on Google Cloud. Alerts go out through Firebase Cloud Messaging.

How they paid for it

Google says Google.org gave Watch Duty 2 million dollars in funding in 2024. It also gave 300,000 dollars in Google Cloud credits. The money paid for compiling satellite, camera and radio data to validate new fires. The 2026 tool is a separate piece of work from that grant.

What a team building a transcription feature can take from this

A general speech model struggles on specialised audio. A fine-tuned deployment on a narrow domain can cut word error rate by almost half. In this case the drop was from 38 percent to 22.34 percent. The cost was six months of embedded engineering work. The training data was 20,000 hours of past recordings. Google's piece says both.

The staffing pattern also matters. Google calls it a team that was "embedded full time" with the customer. That is the forward deployed engineering model: the vendor sends its own engineers to work inside the customer's problem. Buyers of a long transcription deployment can ask a vendor for the same thing, by name.

Google is also the publisher. The figures are its own and the baseline is unnamed. A team picking between Gemini and another speech provider has one data point here. A side-by-side test is still needed.

Source

How Watch Duty is using Gemini to get wildfire alerts out faster, Google blog.

SourceGoogle

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project