# Flock explanation of current Audio Detection architecture, July 2026 > *Audited source extract. The dossier was used only as a lead; the underlying source was independently acquired and checked on 2026-07-20.* ## Source metadata - **Publisher:** Flock Safety - **Original dossier URL:** https://www.flocksafety.com/blog/how-flocks-audio-detection-works - **Resolved/canonical URL used:** https://www.flocksafety.com/blog/how-flocks-audio-detection-works - **Publication date:** 2026-07-10 - **Underlying event or version:** 2026-07-10 - **Archived:** 2026-07-20, from unauthenticated public access; no cookies, tokens, or login state retained - **Wayback snapshot:** save failed on 2026-07-20 after the archive service returned HTTP 403; retry pending ## Verification verdict **supported with narrower wording** — Flock's July 10, 2026 explanation supports the rolling-window, discard/retain, localization, diagnostic-sample, and limited human-distress pilot/removal descriptions as the vendor's current account. It is not an independent audit of classifier behavior. ## Claim boundary This source is authoritative only within the source class and limitations stated above. ## Extract **Extracted-text lines 238-242:** > Most audio is analyzed on the device and immediately discarded, with only encrypted event clips from qualifying incidents retained. > The article addresses common privacy concerns directly, explaining what the system can do, what it cannot do, and the safeguards built into its design. > It also explains why Flock removed its pilot human distress detection feature, reinforcing a commitment to transparency, privacy, and community trust. > ‍ > Before explaining how Flock’s Audio Detection works, it’s helpful to start with what it is designed to do. Audio detection recognizes a limited set of public safety events, such as gunshots, fireworks, car crashes, and sideshows, to provide earlier awareness of critical incidents. **Extracted-text lines 244-248:** > Here is how Flock's Audio Detection (formerly known as Raven) actually works, including the concerns critics have raised and the changes we've made in response. > How do audio detectors work? > Audio detectors continuously analyze sounds in short rolling windows of about five seconds. The audio is processed on the device itself by a dedicated low-power processor that asks a single question: does this sound match a qualifying event, such as a gunshot, vehicle crash, or tires screeching? > If the answer is no, that audio is never recorded, stored, or sent anywhere. > If the answer is yes, the system records, encrypts, and saves only that short audio clip, determines the location of the event, and sends an alert to the appropriate users. **Extracted-text lines 250-254:** > Some critics correctly point out that a system must process audio in order to recognize a sound. That's true. Processing, however, is different from recording. The system briefly analyzes audio on the device to determine whether a qualifying event has occurred. > The change we're making to human voice detection > We previously offered a pilot feature that detected sounds of human distress, including screaming. This feature was designed to help identify potential violent incidents in areas where other public safety tools were less effective. The feature was only available to a small number of customers as part of a limited trial, and was never broadly released. > After careful consideration and community consultation, we decided to remove the feature. > It’s important to understand what the system does and what it does not do. Audio detection is designed to recognize specific sound events, not people. The system cannot understand speech, analyze language, or identify voices. It doesn’t now, and it never has. **Extracted-text lines 255-259:** > The concerns we hear most, answered directly > "The microphones are always on, listening to everything." > The sensors are always analyzing, in the rolling-window way described above. They are not always recording. There is no way for a customer, or for Flock, to open a live audio feed of a location or listen to a place in real time. The system surfaces short event clips, and nothing else. We built audio detection that way on purpose. > "It can record my conversations." > Flock’s Audio Detection is designed to capture qualifying sound events, not conversations. In the rare instance that a qualifying event, such as a gunshot or vehicle crash, occurs while someone is speaking nearby, the short clip may include a few incidental words spoken in a public space immediately before or after the triggering sound at such a volume that it is incidentally captured. The system does not listen for conversations, identify voices, understand speech, transcribe what people say, or allow anyone to search for spoken words. **Extracted-text lines 261-265:** > No acoustic system is perfect, and we don't claim otherwise. The sensors are built to tell fireworks and similar sounds apart from gunfire to reduce false alerts. When the system detects a gunshot or firework, it estimates the location using detections from multiple sensors rather than relying on a single device. Alerts provide information for law enforcement or security professionals to review and determine the appropriate response. They support human decision-making rather than replace it. When the system gets something wrong, those event clips are used as feedback to correct and improve the model. > "An audio alert triggers an automatic dragnet of every nearby car." > A confirmed event can surface the vehicles seen entering or leaving the area around the time of the sound, using nearby Flock license plate reader cameras. That gives investigators a starting point when investigating events like gunshots. It is not an automatic enforcement action, and it doesn't decide anything. A law enforcement officer or security professional reviews the alert and any associated information and decides what, if anything, to do. > "Deleting clips later doesn't matter, because the harm is in capturing them." > While the sensor must briefly process audio to determine whether a qualifying event has occurred, the vast majority of that audio is never retained. Only short clips associated with qualifying events are encrypted, retained, and deleted after the applicable retention period. **Wayback failure**: The bounded Save Page Now pass stopped after the archive service returned HTTP 403 on the first request. This source was not individually retried; retry pending. ## Notes - Tier: 2 — vendor primary/self-description; authoritative only for the vendor's own claims. - The excerpt is bounded to the claims audited here; consult the preserved original for full context. - Cited by: see `citing_pages:` frontmatter; populated after wiki integration.