A collaboration between Lewis McLain & AI

A Balanced Look at Acoustic Gunshot Detection: The Technology, the Track Record, and the Math

If you hear gunshots in Fort Worth, there is a fair chance the police department’s Real Time Crime Center knows about it before anyone dials 911. A July 2026 Fort Worth Star-Telegram report pulled back the curtain on a technology that has been operating quietly in the city for several years: acoustic gunshot detection. Fort Worth’s deployment is a useful case study, because the city has taken a measured, targeted approach at a moment when the technology itself is at the center of a national debate. This assessment looks at how the technology works, where it came from, who sells it, what the evidence says, and what it actually costs per unit of value delivered.

How the Technology Works

Gunshot detection is fundamentally an exercise in acoustic physics plus machine learning classification. A gunshot produces two distinct signatures. The first is the muzzle blast, a low-frequency boom radiating outward from the weapon. The second, for supersonic ammunition, is the ballistic shockwave, the sharp crack of the bullet itself displacing air along its flight path. Detection systems exploit one or both signatures.

There are two basic architectures. The first is networked triangulation. Multiple microphones spread across an area each timestamp the arrival of a candidate sound. Because sound travels at a known speed, the differences in arrival time across three or more sensors allow software to compute the origin point, in the same way a surveyor closes a traverse. Flock Safety’s Raven system, used in Fort Worth, works this way: the city requires three microphones to independently classify the same sound as gunfire before an alert is generated and a computed location is pushed to the Real Time Crime Center.

The second architecture is the self-contained pod. Acoem, Fort Worth’s other vendor, builds units with multiple sensors inside a single housing, allowing each pod to compute a bearing and location on its own. Each Acoem unit in Fort Worth is hard-wired to a camera that automatically slews toward the detected sound, giving dispatchers a live visual within seconds. This makes the pods portable and well suited to special events and temporary hot spots outside the fixed coverage zones.

In both architectures, artificial intelligence does the sorting. Edge processing on the device discards everything that is clearly not of interest, including human voices, and forwards candidate sounds to the cloud for finer classification, such as distinguishing a firework from a gunshot, which are acoustically similar. Flock has stated its aggregate accuracy runs above 90 percent, while acknowledging the figure varies by environment.

The privacy question comes up in every city council chamber where this technology is proposed, and the vendor answers are consistent: the systems are trained to trigger only on impulsive sounds matching gunfire profiles, everyday audio is neither recorded nor retained, and stored clips are short. Acoem records 1.5 seconds per triggering event. Fort Worth retains audio from the last 200 potential gunfire incidents unless flagged as evidence. Skeptics reasonably note that these are policy and design choices rather than physical impossibilities, which is why retention rules and audit provisions belong in every contract.

A Short History

The technology is older than most people assume. ShotSpotter, now operating under the corporate name SoundThinking, deployed its first system in 1997 and built the market essentially alone for two decades. The value proposition rested on a statistic the company has repeated for years and that independent researchers have largely corroborated: only about 20 percent of gunfire incidents are ever reported to 911, and when calls do come in, they are often vague about location and slow to arrive. A detection system alerts within roughly 60 seconds with a location precise enough to search.

By the early 2020s, SoundThinking’s system was in use in more than 180 cities, with pricing generally quoted at $65,000 to $90,000 per square mile per year plus a $10,000 per square mile initiation fee. Detroit’s contract, at approximately $7,000,000 covering 40 square miles, is representative of a large-city commitment.

The competitive landscape changed in 2021 when Flock Safety, which had built a very large installed base of license plate reader cameras, entered the market with Raven. Flock’s pitch is integration: a gunshot alert can automatically activate nearby Falcon license plate cameras, letting investigators draw a virtual perimeter around a shooting scene and pull vehicle evidence within minutes. With more than 2,500 cities already on Flock’s network for license plate reading, Raven often arrives as an add-on to an existing contract rather than a standalone procurement, which has made it attractive to mid-sized and smaller cities. Acoem, a French acoustics firm with deep roots in industrial and environmental noise monitoring, competes with its self-locating pod architecture, and several smaller entrants round out the field.

The Fort Worth Deployment

Fort Worth’s approach is notable for its restraint. Rather than blanketing the city, the police department placed Flock Raven equipment in five areas selected on the basis of historical gunfire data: Northwest 25th Street, Las Vegas Trail, Rosemont, South Riverside Drive at Berry Street, and Stalcup Road at Berry Street. Together the zones cover roughly five square miles. Acoem pods handle special events and emerging hot spots outside the fixed footprint.

Equally notable is the dispatch policy. Between January 1 and November 16, 2025, Raven sensors generated approximately 6,074 gunshot detection alerts. But Fort Worth rarely dispatches officers on an alert alone; confirmation from a 911 caller or a city camera is generally required, with unconfirmed alerts pushed to officers’ in-car computers as situational awareness. A city review found just 22 calls for service tied to detection alerts in that period. Whether that ratio represents admirable discipline or an underused investment is precisely the question every governing body should ask, and the answer depends on what the system is for.

The Success Stories

The strongest documented value of this technology is not crime reduction. It is finding wounded people whom no one called about.

Oakland credited the technology with locating 101 wounded victims in a single year, several of whose lives were saved by response times as fast as two minutes. The most vivid single case is that of Ersie Joyner, a retired Oakland police captain shot 22 times in a 2021 robbery attempt; no bystander initially called 911, but a detection alert brought officers who got him to a hospital in time. In Pittsburgh, a city controller report found dispatch and en-route times roughly 63 percent faster for detection alerts than for 911 calls, and city data identified 13 shooting victims across two years who were located solely because of the system. Detroit’s police chief told his city council in 2026 that the system produced 256 arrests in the prior year and potentially saved the lives of 114 shooting victims, noting that officers frequently arrive to find someone bleeding out whom nobody had reported. On the Flock side, Morgan Hill, California cleared an aggravated assault case that would likely have gone cold by pairing a Raven alert with a license plate search that identified the suspect vehicle within the hour.

The Counterweight

Fairness requires equal attention to the disappointments. The most rigorous independent research, led by criminologist Eric Piza of Northeastern University, studied ShotSpotter in Chicago and Kansas City and found that while the system reliably got officers to scenes faster and increased recovery of ballistic evidence, it did not increase clearance rates for fatal or nonfatal shootings and did not reduce gun violence itself. Chicago’s Inspector General found that alerts rarely produced documented evidence of a gun crime, an investigatory stop, or a recovered firearm. Chicago canceled its contract in 2024 after spending approximately $49,000,000 since 2018, and the city later paid $90,000 to settle litigation alleging that alerts were used as pretext for stops without independent suspicion.

The aftermath in Chicago is itself contested, which tells you how genuinely unsettled this question is. One academic analysis found violent crime continued falling in formerly covered neighborhoods after removal. Meanwhile, a local news organization has tracked 85 cases of gunshot victims enduring long response delays in formerly covered areas because no one called 911, and calculated a fatality rate among those delayed cases nearly three times the citywide average. And notably, Chicago’s 2026 budget includes $5,000,000 that could fund a replacement gunshot detection contract, so even the nation’s highest-profile cancellation may not be permanent.

The honest synthesis: the technology is proven as a victim-location and evidence-recovery tool, unproven as a crime-reduction tool, and its worth in any given city depends heavily on dispatch policy, coverage design, and what officials expect it to accomplish.

The Math: What Does an Alert Cost?

Fort Worth has not published its contract figures in the reporting reviewed here, so the following is an illustrative estimate built from industry benchmarks, and readers should treat it as directional rather than precise.

Using the widely quoted market rate of $65,000 to $90,000 per square mile per year, a five-square-mile fixed deployment would imply an annual cost in the range of $325,000 to $450,000. Flock typically prices below the incumbent, particularly when bundled with existing license plate reader contracts, so the true figure may be lower. Annualizing Fort Worth’s 6,074 alerts over 10.5 months yields roughly 6,940 alerts per year. That works out to approximately $47 to $65 per alert generated.

But raw alerts are the wrong denominator, because multiple sensors alert on the same incident and reloading pauses generate additional alerts. The more meaningful denominators are confirmed dispatches and outcomes. Against the 22 confirmed calls for service in the review period, roughly 25 annualized, the implied cost runs $13,000 to $18,000 per dispatched response. That number looks expensive until it is set against the value of a single located victim who survives, a single homicide cleared through recovered shell casings, or a single wrongful-death exposure avoided. Detroit’s experience, at roughly $175,000 per square mile under its contract and 256 arrests plus 114 potentially saved lives in a year, suggests the per-outcome economics improve substantially at scale in high-gunfire environments.

The metric no vendor publishes, and the one procurement officers should demand, is cost per confirmed gunfire incident and cost per unique outcome: victim located, evidence recovered, arrest made. Any city considering this technology should write those reporting requirements into the contract.

Conclusions

Four findings emerge from the record. First, the physics and the engineering are sound; these systems genuinely detect and locate gunfire quickly and with high accuracy. Second, the humanitarian case, finding unreported shooting victims in time to save them, is the technology’s strongest and best-documented benefit. Third, the crime-fighting case, deterrence and improved clearance rates, remains unproven in the best independent research, and cities that buy the technology expecting lower crime statistics are likely to be disappointed. Fourth, deployment design matters enormously: Fort Worth’s targeted five-square-mile footprint with confirmation-based dispatch represents a fiscally conservative model that captures the life-saving benefit while limiting both cost and the civil-liberties exposure that helped end Chicago’s half-city blanket deployment.

For a technology that listens for the loudest sound a city ever hopes not to hear, the sensible posture is neither evangelism nor rejection. It is the same posture any finance officer brings to any recurring expenditure: define the outcome you are buying, measure it honestly, and renew only if the numbers hold up.

Sources include the Fort Worth Star-Telegram (July 8, 2026), the Fort Worth city manager’s 2025 report, Flock Safety and SoundThinking published materials, City of Pittsburgh Controller reporting, Detroit City Council budget testimony, research by Eric L. Piza (Northeastern University), the Chicago Office of Inspector General, WTTW News, NPR, and CWB Chicago.