DEPTH  ◆  0420M
Ariel Labs  /  Assistive Speech Research

Most assistive speech is fast.
Ours is yours.

For people who can't speak, the existing tools type words and play them back. We're building something different: a system that learns how you want to communicate — your rhythm, your humour, the things you actually mean to say — and gets closer to it every conversation.

Live Session — Demo
Listening
14:02 THEM How was the doctor's appointment?
14:02 YOU Long. They ran every test in the book and still don't have answers. I'm tired.
14:03 THEM I'm sorry. Want to talk about it?
14:03 YOU Not yet. Tell me about your day — I want to think about something else.
14:03 THEM Deal.
SESSION · 0420 LAT · 180MS ● REC
01 The Problem

Today's tools give people a voice. Not their own.

Augmentative and Alternative Communication (AAC) has spent thirty years optimising for one thing: throughput. Get the words out faster. We think that's the wrong target.

What exists now

  • Generic synthetic voices. Robotic, or off-the-shelf, or both.
  • Pictograms and word grids designed for clinics, not conversations.
  • Tools optimised for speed at the expense of style and personality.
  • Static vocabularies. The system never learns who you actually are.
  • Built for compliance and billing. Not for the person using them.

What we're building

  • Responses generated in real time, in your voice, by a model that listens with you.
  • A long-term memory of how you speak, what you reference, what you find funny.
  • Quality and character first. Speed where it doesn't compromise either.
  • A vocabulary that grows with you — the system gets to know you over months.
  • Built for the person using it. Carers and clinicians are second customers, not first.
02 How It Works

A live transcript becomes a real reply, generated for you to pick from.

The other person speaks. Their words are transcribed live. A language model trained on your voice generates a handful of responses you might actually say. You pick one. It's spoken aloud. All in under a second.

SIGNAL PATH LATENCY TARGET: < 1000 MS END-TO-END
STEP 01

Listen.

The microphone captures the other person speaking. Streaming speech-to-text turns it into live text with sub-300ms latency.

STEP 02

Generate.

While they're still talking, the model is already drafting replies — four options, varied in length and tone, written in your voice.

STEP 03

Choose.

You tap the one closest to what you want to say. Or edit it. Or save the whole set as a snapshot and come back to it.

STEP 04

Speak.

High-quality voice synthesis speaks the reply aloud. Every interaction is logged — quietly, privately — so the model gets to know you better.

03 Products

One platform. Two ways in.

The same intelligence powers both. Choose the form factor that fits your life.

Software
PRODUCT / 01

Ariel Web

Browser & tablet — subscription

For people who already have a device. Sign in, set your voice, start talking. Designed to run on an iPad mounted to a wheelchair, a tablet on a desk, or any modern browser.

  • Live transcription and response generation
  • Personal voice library, custom phrases, snapshots
  • Conversation memory across sessions
  • Carer and family sharing controls
In private beta — waitlist open
Hardware
PRODUCT / 02

Ariel One

Dedicated device — one-time purchase

For people who want a purpose-built tool. Bone-simple hardware, touchscreen, dedicated microphone. Designed to work offline as the underlying models mature. Lives on the table, ready to go.

  • All of Ariel Web, in a sealed, reliable device
  • Hardware tuned for noisy real-world conversation
  • Roadmap to full offline operation (2026–27)
  • Direct support & on-site setup for early customers
Prototype operational — waitlist open
04 The Long Game

We don't want to give people a voice prosthetic. We want to build a memory prosthetic.

The body can fail. The mind doesn't. Our job is to make sure the second one isn't trapped inside the first.
Ariel Labs / Position paper, in progress

Every conversation through Ariel is, with consent, a data point. Over months and years, the system learns the things only you would say: the references, the rhythms, the in-jokes, the people you care about, the topics that come back week after week.

That's not voice cloning. That's voice discovery — finding and amplifying the way you already think, so that what comes out of the device sounds more like you over time, not less.

The endpoint isn't an assistant that speaks for you. It's a second mind that knows you well enough to speak with you. We think this is the most important thing assistive technology can be doing in the next decade. We're building it now.

05 Get Early Access

Be among the first to use Ariel.

We're letting a small number of people in over the coming months. People who use AAC, families, clinicians, researchers. Tell us who you are and we'll be in touch.

No spam. No selling. One email when we have something for you.
You're on the list. We'll be in touch.