WHISPER
Who is speaking?
WHISPER is a live AI-mediated language project for conversations between people who do not share a common language.
The practical aim is simple: help a person understand another speaker and answer in a language they barely know or do not know at all — while remaining the visible, audible and responsible participant in the conversation.
The machine may supply language. The human must not disappear from the conversation.
The problem
Translation can already make a sentence understandable. That does not automatically make a conversation yours.
In a live exchange there is no time to stop, open an app, reconstruct context, write a perfect sentence and then return to the person. Meaning, timing, tone and social presence all matter at once.
WHISPER begins from a more practical question:
What can I naturally say to this person right now?
Its narrow interaction model is:
interlocutor speaks → system interprets bounded context → one short reply is proposed → the user speaks
The project is related to Voiceprint, but not identical to it. Voiceprint asks what must survive when a text passes through language and machines. WHISPER brings that question into a live social encounter.
Five assistance contracts
TRANSLATE ME — the human already knows what they want to say; AI transfers it with minimal added authority.
LIVE TRANSCREATION — the human supplies the utterance; AI preserves meaning, intention and style in natural target-language speech.
HELP ME SAY IT — the human supplies the intended meaning; AI helps formulate a natural sentence.
SUGGEST A REPLY — AI proposes a possible answer; the human decides whether it is actually theirs.
WHISPER — bounded live context is used to propose one short immediate response that the human can repeat aloud.
These are not interchangeable features. They represent different degrees of machine authority.
Authorship and agency
The project is not primarily asking whether AI can translate. It asks what happens to authorship when language itself is partly supplied by a machine.
A person may still use their own body and voice while the wording has migrated elsewhere.
At what point does linguistic assistance become delegated authorship?
WHISPER therefore treats latency, misunderstanding, trust, ambiguity, dependence and perceived authorship as part of the research — not merely as interface problems.
Listening without permanent surveillance
A real conversation cannot depend on pressing a record button after the other person has already begun speaking.
The current requirement is therefore continuous listening only inside an explicitly started, bounded session — for example one taxi ride, one meeting or one service interaction.
continuous inside an explicit session ≠ permanent ambient listening
The system must still learn when not to intervene, whose speech matters, whether a reply is actually needed and how to fail safely when the scene is unclear.
Current status
WHISPER is currently an active research and prototyping project, not a finished product.
The working direction begins with low-stakes everyday communication and controlled dogfood testing. Polish is useful as a familiar baseline because partial comprehension allows mistakes to be noticed. A genuinely unfamiliar language provides a cleaner later test of whether the system itself enables participation.
No implementation stack is presented here as final. No public claim of proven latency, accuracy or usefulness is being made before real tests exist.
Prototype first. Record failures. Preserve human agency. Then decide what WHISPER actually is.