The machine guesses once
Speech-to-text takes its first guess as final, even when it is unsure. “Kambal” becomes “camel”, and the request is lost.
言葉の先に、つながる旅。
JAL-PS-02DELHNDAccessible cabin communication
Every passenger,understood.
すべての乗客が、理解される旅を。
Auralink helps passengers and cabin crew understand each other across languages and accents in real time. When speech is unclear, it asks the passenger a simple question instead of guessing.
- Works offline in the cabin
- No app to install
- Hindi, Tamil, Thai, English, Hinglish
The problem
When speech fails at 35,000 feet, nothing repairs it.
Mrs. Sharma, 72, is flying Delhi to Tokyo to see her son. She speaks soft Hinglish. The cabin is loud. She asks for “thoda wotter and kambal”. The crew member asks her to repeat, then guesses. After the second try she stops asking. Later a safety announcement plays in fast English. She nods. Nobody knows she did not understand.
“Please repeat” is the only fix
Saying it again, louder, is hardest for seniors, soft speakers and anyone with hearing or speech difficulty.
A nod is not understanding
Crew cannot tell which passengers actually understood a safety instruction, so they repeat it to everyone.
The real gap is repair: when meaning is lost, in either direction, nothing brings it back.
On board
Hear. Check. Deliver.
A passenger scans the seat QR and speaks. Auralink matches what it heard by how it sounds, checks with her in Hindi, and the crew get one clear card in polite Japanese. The reply comes back in her language.
SEAT 34K · HINDI · LARGE TEXT
Passenger says
“Madam, thoda wotter and kambal…”
Plain speech-to-text: “thoda water and camel”
- Hear
- Check
- Deliver
How Auralink fixes it
Ask, don’t guess. Then make sure everyone understood.
Say it any way
Speak, tap a picture or type, in Hindi, Tamil, Thai, English or a mix. Scan the seat QR code; there is nothing to install.
Listen for meaning
Auralink compares what it heard with words people really use on board, like paani, kambal and dawai, and considers the moment, for example meal time.
Ask the easiest question
If it is not sure, it shows two big pictures to choose from. If it is still unsure, the crew get a “please help” card with its best guesses.
One clear card for crew
The crew see the request in polite Japanese and reply with one tap. The passenger reads the reply in their own language.
Urgent requests skip the queue. If it sounds like a medical problem, the crew are alerted straight away while the passenger confirms.
Try the real engine
Type a request the way a passenger might say it.
This box sends your text to the same matcher the cabin hub uses. Nothing reaches any crew.
Flight phase
The other direction
Know who understood the safety briefing.
The crew send a pre-approved message. Each passenger sees pictures and a caption in their own language, then taps I understood or Explain again.
- Understood
- No answer yet
- Asked to explain again
Instead of repeating the announcement to 54 people, the crew walk to the few seats that need help. At boarding, exit-row passengers get a picture check too, supporting the crew’s own confirmation.
TURBULENCE BRIEFING0 / 54 understood
30ABCDEGHJK31ABCDEGHJK32ABCDEGHJK33ABCDEGHJK34ABCDEGHJK35ABCDEGHJK
Built for a real cabin
Simple for crew. Reliable at altitude.
No internet needed
A laptop-sized hub on board runs the speech engine. Phones connect over the cabin network. If the hub fails, crew simply carry on as they do today.
Scan and speak
The seat QR code opens Auralink in the browser, already set to that seat.
New language? Add words
Languages are word lists in a file. No retraining: Tamil and Thai were added this way.
One tap for crew
A card, an Acknowledge button and polite replies in Japanese.
No phone? No problem
A tactile seat panel with four raised buttons, or the in-seat screen, does the same job.
Dignity and privacy first
Large text for seniors, a picture-only mode for hearing or speech difficulty, gentle wording like “Let’s find it together”, and speech used only for the request, never stored.
Measured, not promised
Our first numbers, on a laptop CPU.
0.00 sspeech to text, English and Hindi at once
≈ 0.0 sfrom the end of speech to the confirm screen (target under 2 s)
0%of requests understood on synthetic test voices with cabin-like noise
First-try accuracy, same 65 accented and code-mixed requests
Baseline: first guess, keywords56.1%
Auralink, hybrid matcher91.2%
A development benchmark the matcher was tuned on, so it is optimistic; wrong confident cards: 0% for both. Final claims come from held-out volunteer recordings.
The 88% comes from computer-generated Indian English and Hindi voices, so treat it as a smoke test. We are recording real volunteers, including seniors, to measure accuracy, wrong-action rate and recovery properly.
Live demo
Open the cabin.
Checking the hub…
First time? Start here: who uses which screen, the seat QR code, the crew PIN, and how it works with no internet.