How FireFeed reads the radio
Five steps from a transmission to an incident. The demo enters at the second step with text; the rest runs as it would on your own recordings.
- Radio callaudio or a transcript
- Transcribeskipped when text exists
- Analyzestrict JSON schema
- Groupcalls into incidents
- Feedthis site
Input
One file per transmission: a transcript (.txt) or audio (.wav, .mp3, .m4a, .ogg, .flac, .webm), plus a JSON sidecar with an ISO 8601 timestamp that carries a UTC offset, and the channel name. Ingest checks every file before the first request and reports all problems at once.
Transcription
Audio goes to OpenAI speech-to-text (gpt-transcribe in firefeed.toml), cached by the SHA-256 of the audio bytes, model, language and prompt. A transmission that already has a transcript skips this step. This demo is text-only, so no transcription ran and the site has no audio player.
Structured extraction
Each transcript goes alone to the Responses API (gpt-6-luna, reasoning effort low, store: false) with Structured Outputs: a strict JSON schema, seven required fields, no extra keys. The answer is then validated locally against the same schema and against rules the schema cannot express, such as not_incident if and only if the status is none. A failed check gets one retry; after a second failure the call is marked as failed.
| Field | Type | Meaning |
|---|---|---|
is_fire | boolean | part of the response to a fire call, true even when the crew finds no fire |
incident_type | enum of 12 | structure_fire, vehicle_fire, wildland_fire, mva_with_fire, … unknown, not_incident |
location_as_heard | string | null | the incident location as spoken, including a command name such as “Maple Command” |
units | string[] | every fire unit named, as type and number |
status | enum of 8 | dispatched, en_route, on_scene, working, under_control, cleared, cancelled, none |
summary | string | one plain sentence |
confidence | number | 0 to 1, checked locally |
Local knowledge
The instructions carry the county gazetteer: 20 street names, 8 unit names, the dispatcher's call sign and local 10-codes (10-97 arrived on scene, 10-8 back in service). After extraction, location_as_heard is matched to a street: exact words first, then a fuzzy comparison over joined words, with digits required to match. “Bell weather” becomes Bellwether Avenue and “Thimble Crick Road” becomes Thimble Creek Road, while “County Road 21” would not match County Road 12.
Grouping
Calls about an incident are grouped by rules, applied in order. Two different known streets never join. A new dispatch joins only an incident at the same street. The same street links strongest, then a shared unit; a call with neither joins the incident heard on the same channel within 120 seconds. A cleared or cancelled call closes its incident, and an incident silent for 60 minutes takes no new calls. Every threshold lives in firefeed.toml.
Cost control and measured cost
Every answer is cached under a SHA-256 of the model, settings, instructions, schema and input, so a repeat run makes no requests; the committed cache lets anyone reproduce this run without a key. Before a batch, a dry run counts cache misses and estimates tokens and cost. The run refuses to start above a request cap and asks for confirmation above a cost threshold.
| API requests | 53 |
|---|---|
| Input tokens | 55,454 (1,046 per call) |
| Output tokens | 6,565, of them 2,498 reasoning (124 per call) |
| Cost | $0.0088 in total, $0.00017 per call |
Prices: $0.10 per million input tokens, $0.01 cached, $0.50 output, Standard tier, from the OpenAI pricing page, checked 2026-11-21. Output tokens include reasoning tokens.
Accuracy on 53 synthetic calls
The expected answers were written by hand from each transcript alone, before the run, and were not changed after it. 47 of 53 calls match on every field.
| Field | Correct |
|---|---|
| Fire call or not | 53 of 53 |
| Incident type | 50 of 53 |
| Street | 52 of 53 |
| Units | 53 of 53 |
| Status | 50 of 53 |
| Incidents grouped exactly | 13 of 13 |
Every difference goes the same way: the model is more cautious than the labels.
- Status, 3 calls (c006, c020, c040): arrival reports that already name the first action. Labeled
on_scene, answeredworking; the labeling rules do not say which stage wins. - Incident type, 3 calls (c007, c009, c026): the type follows from context but is never stated. Answered
unknown. - Street, 1 call (c007): “the hydrant at Bellwether and Tinsmith” was labeled as the incident street; the model returned no location. A hydrant is not the fire's address, so the label is arguable.
Labeling rules and the full comparison: data/README.md in the repository.
Limits
The transmissions are synthetic and cleaner than real radio, which has crosstalk and clipped words. Accuracy on real traffic has not been measured. A real channel can be connected only with the written permission of the source's owner.
Stack: a Python 3.12 pipeline (openai, jsonschema) covered by pytest and ruff; this site is a static Next.js export with no backend and no API keys.