Vigilant Ear
EnglishEspañolPortuguêsFrançaisDeutschItalianoالعربية日本語简体中文繁體中文한국어Русскийहिन्दीTürkçeRomână

Vigilant Ear 👂🛡️

Effective as of version 1.1.12 · October 2026.

An acoustic radar for people who can't hear.

An app built specifically for the Deaf, hard-of-hearing, and CODA community. Most sound-recognition apps tell you what a sound is. Vigilant Ear tells you where it is, who's making it, and what they're saying — turning an iPhone into a device that describes the sound around you as it happens.

A siren's direction and distance. A knock behind you. The people in a conversation, drawn as separate transcribed voices — each one captioned and directionally placed. If someone is speaking a language you don't read, their words can arrive translated into yours. Alerts reach your Lock Screen, Dynamic Island, and Apple Watch so a glance is enough.

Everything that matters runs on the device. Audio is not recorded or uploaded for recognition. Nothing depends on a human ear hearing anything.

Who it's for

Vigilant Ear is an accessibility aid, not a certified life-safety device.

What it does

🧭 It sees sound — direction & distance

Using the iPhone's two microphones, Vigilant Ear measures the angle of a sound and whether it is ahead of you or behind you, holds that reading steady to about a degree, and places it as a live marker on a heading-up radar ring and map. Two microphones on one line can't tell left from right on their own. A sound 50° to your right and one 50° to your left arrive with the same tiny time gap. The app draws the reading plus a fainter ghost on the other side, and a turn of the wrist or a second phone settles it. Distance is estimated from loudness against a street-calibrated curve and shown as the estimate it is, with a likely range, in the units of the country you are in rather than the language you read the app in. On a US phone that reads ≈ 50 ft. Move, and the markers hold their real-world position. This is the core: spatial awareness of a world you can't hear. (Numbers below.)

🚨 It recognizes important sounds — and warns you

An on-device classifier identifies hundreds of everyday sounds and watches the critical categories — sirens, alarms — including a dedicated car-alarm class — doorbells/knocks, baby cry, a person nearby, and severe weather. When one fires, you get a clear on-screen alert, optional push notification, and a distinct haptic — even when the app is backgrounded or the phone is asleep. Critical categories default ready so enabling notifications doesn't mean “everything off.” Turn all alert categories off and the engine fully hibernates while backgrounded to save battery. A Sentinel layer cross-checks alerts against independent evidence — direction, motion, and public feeds — so what fires is corroborated, not a lone classifier guess. It works both ways: a siren-shaped moment inside a song gets held, but a real siren repeating through your music breaks through and alerts.

Severe-weather warnings come from official government sources in more than 140 countries and territories, with more added often — free for all users. Civil-defense emergency alerts are included too, with their own Emergency Alerts toggle. Feeds are narrowed to the ones that cover where you are. Where available, the app also reads national weather agencies' advance bulletins — the plain-language notices issued days before a typhoon or heavy rain, not only the warnings issued once danger has arrived — plus the sudden-downpour and landslide bulletins. Advance notices are shown quietly, without the sound and vibration reserved for a real warning.

🌧️ Weather radar on the map

Rain and snow show on the map from the moment the app opens — from the national radars of the United States, Canada, Japan, Italy and much of Europe, and a satellite rain estimate across most of the rest of the world. Touch the map and the last hour plays back three times, so you can see which way a storm is moving; a scale at the bottom reads from Light to Dense. If you don't want it for a session, tap the weather button to turn it off. The radar arrives the same way the warnings do — through our own server, so your phone never contacts government servers. Light grey arrows also show which way the wind is blowing, anywhere in the world, from NOAA's global forecast for the current hour.

🛟 Who to call — the life ring

A life ring beside the map's compass opens a short list for the country you are standing in:

⌚ Apple Watch + Live Activity — glance and know

💬 Speaker Mode — live, directional captions (free)

Turn on Speaker Mode and Vigilant Ear transcribes people talking near you into caption blocks, one per voice. On-device speaker diarization keeps voices distinct — who is saying what — with a directional cue on the inner ring. The live speaker is highlighted; older text scrolls away as space is needed.

Each speaker in their own language. On iPhones with at least 8 GB of memory, a one-time captioning-model download (about 1 GB — over Wi-Fi, or on cellular after asking) lets the app caption each turn in the language actually spoken, so a room that switches between English, Spanish, French and Arabic is understandable. In our tests, word errors in a four-language conversation fell from about 80% to 14%. Since version 1.1.12 an improved speech model makes far fewer mistakes in noise — in a loud bar on an iPhone 17, German captions went from 45% of words wrong to 20% — and replaces the old one by itself over Wi-Fi. Without the download — and for readers of Chinese, Japanese and Korean, where Apple's recognizer reads better — captions use Apple's on-device speech recognition as before.

Two things most caption apps won't do.

Honesty about confidence. A subtle per-row quality mark and dotted underlines beneath doubtful words tell you when to trust a line and when to double-check.

Self-correction. Right after a sentence lands, the app re-reads the audio with full context and can restore a missed or misheard word within a couple of seconds, then the text freezes.

Name Called (Preferences, off until you turn it on) watches those captions for names you list — yours, a child’s, a partner’s. When someone says “order for Marie ready,” you get a distinct haptic and Watch alert with direction, not another caption bubble. The words still appear in Speaker Mode as speech; this is the tap that they were talking to you.

Strong language can be masked (Preferences → Captions). Swearing otherwise arrives as plain text with nothing to soften it, and a reader has no chance to look away first; turn this on and those words are replaced with symbols, so the sentence still reads without the word. Your device's own speech recognizer decides what counts — not a word list of ours — so it follows the language being transcribed. On an account declared as under 18 it is always on and shows a lock. It applies to speech this device transcribes; text relayed from a paired phone was transcribed there and arrives as finished text.

The word a speaker leaned on is marked in bold. Said aloud in English, "I didn't say THAT" and "I didn't say that" are two different sentences, and an identical transcript loses the difference entirely. The app watches for the one word a speaker's voice lifted in pitch and sets it in bold — at most one per line, and nothing at all when it isn't sure, because marking the wrong word would claim the person said something they did not. Off in Chinese and Japanese, where pitch tells you which word it is rather than how hard it was pushed.

Captions are free; automatic translation is the optional Power Pack+ layer. Captions can also be spoken aloud to a Bluetooth device, including a hearing aid — free, in Preferences. Direction Tones (also free, in Preferences → Captions) add an optional audio cue in the ear you choose that signals where a speaker is — useful with a hearing aid or single-sided hearing, and available to anyone.

🌐 Auto-Translate — your language, live (Power Pack+)

With Speaker Mode on, when a nearby person speaks another language, Vigilant Ear can detect it and render their captions in your language, with the source language shown on their block. The chain — hear → separate speakers → transcribe → translate → display — runs on the device; the only network moment is a one-time language-pack download from Apple, and when that download is what you are waiting on the app says so rather than leaving you looking at untranslated text. You don't have to know or pick the other language first. A finished sentence is translated as soon as it ends, without waiting for the speaker to stop, and translations arrive in a fraction of the time.

This is the closest thing available to science fiction's universal translator — a device that simply understands. Vigilant Ear detects the language on its own, follows every speaker in the room, and captions them all in your language — no earbuds, no setup, on your device.

graph LR
    A["Just listen"] --> B["Language detected automatically"] --> C["Every voice separated"] --> D["Whole room captioned in your language"]

🎵 Music & broadcast awareness (Power Pack+)

ShazamKit identifies music playing around you and detects song changes automatically so the title stays current.

🎛️ Acoustic Scope — see sound like an engineer

A professional live view of the sound around you: spectrum, spectrogram, ⅓-octave RTA bands, chroma, and harmonic partials — free for everyone. Partials reads music: each tone is named by its note, with its harmonic series beside it, higher sounds such as smoke alarms are named at their true pitch, and you can set a target note that draws as a line to sing or tune against. Leaving the Spectrogram open costs next to nothing, however long you watch. The tools that capture sounds for training your own custom sound packs are part of Power Pack+.

📦 Custom Sound Packs — teach it your world (Power Pack+)

You can teach Vigilant Ear the sounds that matter to you — local birds, a building's door chime, a generator's abnormal hum, or a type of plane taking off. These add-on packs stack on top of the built-in detection, so new sounds never crowd out sirens and alarms. A step-by-step guide is built into the app.

🛰️ Constellation — many iPhones, one shared ear (Power Pack+)

With two or more Ultra-Wideband-enabled iPhones (most since iPhone 11), Constellation mesh pairs them so they can sense each other's position and fuse what each one hears into a single, more precise picture of where a sound is coming from — a distributed, passive listening array. Captions are fused too: each phone transcribes what its own microphone hears, and the words are aligned across phones so the one nearest a speaker contributes what it heard — words only one phone caught are kept rather than lost.

The phones connect directly to each other — no router, no shared Wi-Fi network and no internet; Wi-Fi simply has to be switched on. They use the same peer-to-peer link AirDrop uses, and only phones with the right hardware can join. The link between the phones is always encrypted; phones running Vigilant Ear 1.1.6 or older need to update the app to participate in the mesh.

Partner messages — send a quick text to a mesh partner's phone; it lands in their caption feed and can arrive translated into their language, on-device. Messaging is age-aware: built on Apple's Declared Age Range, texting between an adult and a minor stays off unless both people have deliberately named each other. Alerts and captions are never gated — only person-to-person texting.

🔗 Remote Link — reach someone who isn't with you (Power Pack+ to start)

The thing a phone call would normally do, done with video and text instead. You send an invitation code; the other person joins from inside Vigilant Ear without owning Power Pack+. Unlike Constellation there is no proximity requirement — the two of you can be anywhere.

No audio is used at any point. Video and text only, so nothing about the link depends on hearing at either end — and it gives you a way to sign with someone through the app. The app itself does not understand sign language; it carries the video and the two of you do the rest. The connection is direct between the two phones wherever the network allows it, and the app shows you plainly whether it is Direct or Relayed, so you always know if your video is passing through a relay.

Captions travel too. What each phone transcribes appears on the other, translated into the reader's language when it needs to be. Either of you can pause the link — video and captions — and resume it, and hang up from the main screen.

📷 Camera AR — “see the sound”

Open Camera AR from the Action Fan and pin detected sounds at their real bearing in the live camera view. Markers cluster by speaker or by sound category and direction so the view stays readable, and they fade as a sound stays quiet.

🗺️ Maps, roads & path prediction

Sound bearings project onto real GPS coordinates on the map. Vehicle sounds can be snapped to nearby streets and their paths predicted so a passing truck reads as moving along the road rather than through buildings. (Try the fire-truck demo in the Feature Playground.)

🪄 Feature Playground — prove it without ears

Feature Playground is public for everyone: Home & Street sample alerts (knock, alarm, baby, siren, weather), multi-phone and conversation demos. These have a clear watermark so practice never pretends to be a live event. Closing the panel tears demos down cleanly (no stuck GPS spoof, no leftover flags).

♿ Accessibility first

Built for Deaf / hard-of-hearing / CODA and color-blind users:

Free & Power Pack+

The safety core is free, forever:

Power Pack+ is a one-time unlock (not a subscription) with a 90-day free trial. It adds the superpowers:

Free or Power Pack+, your audio stays on the device for recognition — the tier only changes which features are unlocked, never where raw audio is sent for analysis.

How it works (under the hood)

Vigilant Ear is a local-first, on-device pipeline built in layers: capture once, then let independent specialists read the same sound and check each other's work. Raw audio is captured on a high-priority tap, copied into a pooled buffer free-list (no alloc thrash on the realtime path), and fanned out without stalling the UI:

graph TD
    A["Stereo mic tap"] --> B["Pooled buffer snapshot"]
    B --> C["Sound classifier<br/>(Apple Neural Engine)"]
    B --> Y["Second-opinion classifier<br/>(YAMNet, ANE)"]
    C --> S["Sentinel — evidence layer<br/>corroborates, vetoes, escalates"]
    Y --> S
    S --> H["Alerts · haptics · Watch · Live Activity"]
    B --> D["Spatial math<br/>TDOA · level trend → bearing · distance · approach"]
    D --> R["Radar ring · map · Camera AR"]
    B --> F["Speech recognition<br/>(SpeechAnalyzer, or per-turn models)"]
    B --> N["Turn boundaries<br/>(Sortformer diarizer, ANE)"]
    N --> E["Voice identity<br/>(ReDimNet embeddings, ANE)"]
    B --> E
    F --> G["Captions — one per voice"]
    E --> G
    G --> T["On-device translation<br/>→ your language"]

No single model is trusted alone. The Sentinel layer sits between detection and alerting: independent engines corroborate or veto each other frame by frame — and when repeated real-world evidence piles up against a veto (a real siren wailing through your music), the evidence wins and the alert fires.

Captions get their own second chance. Every finalized sentence is quietly re-read from a short in-memory audio ring with full context — corrections land within about two seconds, then the text freezes for good:

graph LR
    L["Live caption appears"] --> K["Raw audio re-read<br/>with full context"] --> V["Guarded comparison<br/>(keeps what was heard)"] --> W["Row quietly corrected<br/>≤ 2 s, then frozen"]

Weather and earthquake alerts take the opposite path from audio — nothing about your sound ever goes out, but alert data comes in. Every official feed — and the radar and wind images — flows through a small relay cache we operate, so one fetch of the public data serves every user, and your phone never contacts government servers:

graph LR
    P1["Vigilant Ear<br/>On Your Device"] --> W["Wingdings alert cache<br/>one shared copy · 5-minute refresh"]
    W --> N["Official public feeds<br/>weather sources in 140+ countries and territories<br/>NZ civil defence · earthquakes"]

What your phone can hear — measured

Your iPhone has two microphones about six inches (15 cm) apart, a barometer, and a very good clock. That is enough to tell you which way a sound came from, roughly how far, whether it is coming toward you, and whether a door just opened. Here is how accurate each of those is, how we checked, and why it matters if you can't hear the sound yourself. The direction table was measured on an iPhone 17 in September 2026. The same desk check on an iPhone 16 Pro Max did not yield a continuous angle — that image snaps a sound toward ahead or beside — so the degree rows are the 17's alone. Distance stays an estimate by nature — we say so on screen. The taped distances in this section are in feet, the units of that tape.

Direction — to a couple of degrees. A sound reaches the top microphone before the bottom one, or after, and from that gap the app works out how far off the phone's axis the sound is and whether it is in front or behind. The phone's two channels are not plain microphones but a processed stereo image, and that image adds a fixed pattern of its own whenever a room is noisy. The app learns that pattern from the quiet after launch. After about a minute of a quiet room it takes the pattern out before reading the direction; until then the reading runs without that correction.

What we measured Result
Median error across 96 simulated rooms, quiet office to hard-walled room at 5 dB SNR 1.0°
iPhone 17 on a desk, room noise on, speaker 30° off the axis read 61.7° from the side plane against an expected 60°, wobble 1.6°
Same, speaker straight off the phone's side read 4.5° against an expected 0°, wobble 2.2°
Same, speaker 60° off the axis median 35.7° against an expected 30°; the run split, with a second cluster at the straight-ahead delay, most likely a reflection off the rig
Same, speaker straight ahead every reading at the maximum delay, as it should be
Readings landing on the image's own pattern instead of the sound none, at any angle

Why it matters: if you can't hear a siren, "behind you, off to one side" tells you where to look. A reading that holds still is only worth trusting once it has been checked against a tape measure, and this one was checked twice. The first check made the result look better than it was.

Distance — shown as the estimate it is. One microphone can't measure distance, only loudness, and a loud truck far away can sound like a quiet car up close. The app estimates distance from loudness using a curve fitted on a real street, then says it the way a careful person would. On a US phone that is about 50 ft, likely between 25 and 100.

What we measured Result
Real car detections replayed through the calibration 1,131
Landing inside the 100-foot road they were actually on 73%
Far-lane cars estimated 78 and 87 ft where the tape said 73 and 85 ft

The road (a city intersection) was measured lane by lane; when the curve errs, it errs on the near side. The calibration is locked by an automated test so it can't silently drift. Alarms inside your own space, like a smoke detector, get no distance figure at all — they are already in the room with you.

Motion — coming toward you. For cars and sirens, getting louder usually means getting closer. The app watches how fast a sound's level rises over a few seconds: a sustained climb of about 1.5 dB per second (a clear, steady increase in loudness) means approaching, a matching fall means leaving, and a sound that stops changing stops being called either after about four seconds. A single loud moment — a horn tap, a door slam — can't trigger it. Built on physics and nine automated tests; a recorded street pass-by is the next field confirmation.

The air in the room — a door has a signature. The barometer can notice a door opening about fourteen feet (4.3 m) away: pressure dips as the door swings, then rebounds as it closes. Over fourteen hours of quiet and fifteen door events, every door made that two-part shape and nothing else in the house did.

Event Pressure rate Shape
Front door, opened and closed, ×10 0.6–1.8 Pa/s dip, then rebound, ~5 s apart
Front door, slammed, ×5 1.1–2.5 Pa/s same pair, taller
Air conditioning cycling overnight, ×8 0.11–0.46 Pa/s slow ramp over minutes, identical on both phones
A quiet room ~0.02 Pa/s nothing
Human activity, walking by — the motion sensor notices and the phone ignores it

A door that doesn't seal the room was invisible — a screen door, in this test — exactly as it should be. Today a pressure shift can show as deep rumbles on the map; a dedicated quiet-night door alert is measured and designed, and is not yet the everyday default.

Constellation — two phones settle the side. Each phone on the mesh draws a line toward the sound from where it sits, and the lines only cross cleanly on one side. When they do, the ghost disappears and the app can say left or right again. In simulation, two phones 40 meters apart place a siren 50 meters away to within about 2 meters.

How anything here gets believed. Every number went through the same three gates, in order: a simulated room with the answer known in advance; small automated tests that recreate the exact case that would fool the app and run on every change; then the real phones on a real desk with distances measured by hand. Nothing counts as accurate until it survives the third. For someone who can't hear the sound, the app's word is the only word — it should be earned the way a lab earns it.

Deeper notes for engineers: Physics — TDOA, distance, motion, Constellation, and barometric detection, with every figure tagged MEASURED / BENCHED / MODELLED.

Privacy

Full details: PRIVACY.md · TERMS.md · SUPPORT.md

Hardware & platforms

Localization

Fully localized — interface, alerts, and captions — into English, Spanish, Portuguese (Brazil), French, German, Italian, Turkish, Arabic, Japanese, Simplified Chinese, Traditional Chinese, Korean, Russian, Hindi, and Romanian (15 languages). Follows the system locale or a manual choice in the app. Romanian captions work on-device; Apple Translate has no Romanian, so Auto-Translate shows the original for that language.

Status & disclaimer

Vigilant Ear is an experimental acoustic-accessibility aid, not a certified life-safety utility. Localization resolution varies with surroundings, weather, wind, and microphone hardware. Always maintain your typical environmental awareness — don't rely on it as your only source of safety information.

Some capabilities (camera AR markers, Critical Alerts entitlement upgrade when granted by Apple, advanced multi-pack sound authoring) continue to evolve; the free Home / Street watch and live captions are the product you can trust on day one.

Sources and Licenses: api.vigilantear.com/sources

Contact: vigilantear@wingdingssocial.com

Made with ❤️ for the D/HH community and acoustic research.

Wingdings, Inc.

© 2026 Wingdings, Inc.
All rights reserved.
Multiple Patents Pending