You are building an app
Stop rebuilding turn-taking, reading levels, translation and memory. Call the API, render the choices your own way, and spend your time on the part only you can build. Start with the API.
TinkySpeak, as an API anyone can build on.
Millions of people speak through a board of words and symbols rather than a voice. TinkySpeak is the engine behind that conversation — every feature of it — now available as a runtime API. You bring your symbols, your voice and your interface; we bring the conversation.
What it actually does.
Someone says something to the person you are supporting. You send us that sentence. You get back complete, speakable replies in their words, at their reading level, in their language — and the thread of the conversation is remembered between turns.
you → 'Would you like tea or coffee?'
us → six complete replies, each with a label, a sentence and an id
you → the person taps one in YOUR interface
us → the exact words to say, and the language to say them in
you → your voice engine speaks it, and you tell us it was saidThree ways in.
You teach or work in a clinic
The same engine runs behind whatever app your team prefers, so a board can follow a student between devices and languages. Try the live board, then tell us what your classroom needs.
You run a community or a family project
Small projects are exactly who this is for — most AAC apps start as one family's need. Ask for a key, use your own symbols, and share what you make. Write to us.
What comes with the key.
Complete replies, never fragments
Six by default, one to twenty-four as you choose. Each one is a whole sentence with a label, an emoji and an id you select by.
Memory that counts up
Both sides of the conversation, kept turn by turn until you reset it. Only delivered words enter the model's context.
Two languages at once
The partner's words and the person's words, each kept in their own language and translated for delivery. BCP 47, regional variants included.
A camera when you want one
Send a photo and the replies become about what is in it — a menu with its prices, a shelf, an object, a room.
Reading level that fits the person
One symbol per tile up to forty-word sentences. A reply that does not fit is discarded, never cut short into something else.
Your model or ours
Gemini, an Ollama model you run, any compatible endpoint, a board you wrote by hand, or a local model. Online or offline is your call.
Because the hard part should only be built once.
Every AAC team solves the same conversational problem before they can build the thing that makes their app worth using. That duplicated work is why good ideas in this field stay small. Hand it to them as an API and a two-person team can ship what used to need a company.
The rule that governs all of it
A communication system cannot afford to invent. When a reply does not fit the person, it is discarded rather than truncated into a different meaning; when a translation fails it is reported rather than guessed; when the evidence is not there the answer is “unknown”. Putting words in someone’s mouth is the one failure this field cannot tolerate.
ContinueThe API