Why Vigilant Ear?

Updated October 2026.

This isn't just any sound-detection app. Those recognize a short list of noises and put an icon on your screen. Most of what you actually miss isn't on that list.

Someone behind you says your name. You don't turn around. They decide you're rude.

A barista calls your order. A nurse calls your name in the waiting room. You don't answer, and they all draw the same conclusion.

Your phone heard every one of them. It just had no way to tell you.

Vigilant Ear changes that — on the phone already in your pocket.

Any language — it works out what's being spoken; you never pick one
Who said it — each voice in its own color
Where it came from — a bearing and a distance on a live map
How sure it is — a confidence mark on every line
When it's you — a distinct tap when someone says your name
Private — your audio never leaves your phone
No subscription — captions and alerts free forever

And it isn't one problem — it's three. A doorbell is the easy case: loud, sharp, over in a second. Traffic building on the road outside is the hard one — far off, broadband, no clean edge to grab hold of. An earthquake is the case nobody hears at all: it arrives as motion and pressure below the range of human hearing. Your phone's barometer and accelerometer feel it.

Most assistive apps handle the first. Vigilant Ear handles all three simultaneously.

The part nobody puts in the brochure

Every listening app leads with the smoke alarm. It matters, and Vigilant Ear catches it, but it's the easy part: loud, rare, unmistakable. What slips past you is the subtle stuff. A stressed word. A quiet person walking nearby at night. A baby's cry in another room.

Standing Watch is a lamp on your screen that's always on: steady cyan while the room holds its usual pattern, amber the moment something changes — a new voice, a sudden quiet, something approaching.

Name Called watches for the names you choose — yours, your kid's, your partner's — and gives a distinct tap when someone says one. Your Apple Watch and the Standing Watch lamp both show which way it came from. The sentence still appears in captions like any other speech. The alert isn't someone spoke. It's someone is talking to you.

Where it came from

A siren isn't information until you know which way it's coming from. Every sound Vigilant Ear hears goes onto a live map with a bearing and a rough distance, so an alert answers the question you actually have: do I need to move, and which way?

Direction Tones can send that bearing to one ear, for single-sided hearing.

Who's speaking, and how much to trust it

Live captions separate the people around you by voice, each in their own color, so a conversation reads as a conversation instead of one unbroken block.

Most captioning apps hand you every word with the same quiet confidence and leave you to find out later which ones were wrong. Every line here carries a three-bar mark — three bars means it heard clearly, one means take it with a grain of salt — and the individual words it wasn't sure of are underlined. In a noisy room, knowing which words to trust matters as much as the words.

When someone leans on a word — I didn't say he took it — that word is highlighted, so the stress a transcript usually throws away still reaches you.

You never pick a language first. Vigilant Ear works out what's being spoken and captions it. A room with three languages in the air is still a room you can follow.

It can translate almost all of them into your language. Where it can't, it tells you so and shows you the original, rather than handing you a guess dressed up as a translation.

One phone points. Two triangulate.

One phone's microphones sit a few inches apart, which is enough to tell you a direction. Put a second phone across the room and the array is as wide as the room: the bearings from the two phones cross, and a direction becomes a position. Captions merge instead of competing, and phones that have never met form that array on the spot, working out to the foot where each one is standing. Nobody has to be in the same conversation for it to help.

Remote Link reaches someone who isn't with you: video and caption text, with no audio track at any point — so nothing about it depends on hearing at either end, and it gives you a way to sign. Your captions cross as text and arrive in the other reader's language. You send a code; they join from inside the app without buying anything.

It's encrypted end to end between the two phones. The relay that forwards it can't read it, we can't read it, and nothing is recorded — not on either phone, not on any server. The full path.

Warnings from the people who issue them

Severe weather comes from official government sources in more than 140 countries and territories — with more added often — filtered to where you actually are. Earthquake alerts wait for official confirmation rather than guessing.

Those warnings reach you through our own service rather than your phone calling each agency directly. So your phone never contacts a foreign government's servers, and a busy public feed can't drop your alert because too many people asked at once. Your phone only tells us roughly where you are — a square about 50 km across — and checks whether you're inside each warning area itself, so your exact location never leaves it.

Several small experts, not one big one

The obvious way to build a listening app is to run one model and report whatever it says. Vigilant Ear runs a team of specialists instead, each doing one job it's good at. One recognizes everyday sounds, and a second checks its work. One recognizes which language is being spoken, out of more than a hundred, and hands the speech to the caption engine that measured best for that language. One separates the voices in a room, and another keeps track of which is which. One decides whether music is really playing. One masks strong language when that's turned on. Where a sound came from, and how far, is worked out from the microphones themselves.

They disagree often, and that's the point. A siren-shaped moment inside a song looks exactly like a siren if you only ask one model, so a siren alert has to get past a second opinion (is that music?) and hold for more than a moment before it interrupts you.

It also lets the app say it's unsure instead of guessing: the confidence marks on captions, the voice left unattributed rather than given to the wrong person, the alert that waits one more frame for corroboration. A single model can't know when it's wrong. Several can, as long as something decides, moment by moment, which one to believe.

We didn't set out to file patents. We set out to make this work in a real room — loud, crowded, in more than one language at once — and that kept needing pieces we couldn't find anywhere. So we built them, and a few are now patent pending.

The Acoustic Scope puts that raw perception on screen — spectrum, spectrogram, ⅓-octave bands, chroma, harmonic partials. We tuned its levels on real phones against a reference sound level meter, so an engineer can work with the numbers. It's free for everyone, because for a lot of people, watching the shape of a sound is the fastest way to understand it. There's a field guide to reading it.

Its Visualizer turns the room's music into a light show: every beat detonates a ring, the bass makes the whole scene breathe, the highs rain down as sparkles. Put it on a TV or a projector and the wall joins the party, with your DJ name in the corner like the artist's. You don't have to hear to entertain a room.

Private by design

Listening, transcription, translation and sound classification all run on your phone. The work that touches what you say and what you hear isn't something we can read, because it never reaches us. That's a deliberate constraint, not a flourish: models small enough to live in your pocket never need to send what they hear anywhere.

A few things do use the network: maps; the weather and earthquake alerts above; road context; Music ID, which sends a short audio fingerprint and never recorded audio; the optional Caption Support models, a one-time download; the Research Array, which asks you once and then shares low-frequency and ground-motion readings for earthquake research — when, and roughly where, to about 1 km, never audio; App Store purchases; and Remote Link, which travels between the two phones by definition — encrypted end to end, so nothing in between can read it.

No ads. Nothing sold. No account required.

What it costs

Free to try at first. After the trial, captions and alerts stay free forever — the safety floor isn't what we sell.

A one-time payment keeps Power Pack+ for good: Auto-Translate, which turns a language you don't read into one you do, as it's spoken; Music ID, which names the song in the room so you know what everyone else is hearing; and hosting a Constellation or a Remote Link — the people you invite join free. No subscription, ever — and no new hardware to buy.

Vigilant Ear is an experimental research and accessibility aid — not a certified or life-saving device. Always use your own situational awareness.