Generating speech is finished work. A model can read a paragraph in a voice you would not question. Ask that same model to hold a conversation and it comes apart somewhere else entirely: it talks over you, it answers half a second too late, it treats the pause where you were thinking as the end of your sentence, it replies to the question you abandoned mid-way. None of those are audio-quality problems. They are interaction problems — timing, overlap, silence, interruption, who is speaking to whom, and what a person actually meant by stopping. That is the layer EESI builds.
The next generation of AI will not be defined by how well it generates speech, but by how well it understands and navigates interaction itself. From AI that responds to AI that participates.
Three things make that possible, and you can see all three in the API. A representation of conversation that keeps what a transcript throws away. Simulation, where models talk to models until something breaks. And conversation-native models trained on continuous audio instead of on text scraped out of it.

​
Where interaction shows

These are the surfaces where the difference is audible. Start here, not with the file endpoints.

Live

Speech in, speech out, one model, one socket. Barge-in, end-of-turn detection and cancellation are handled by the server, not by you. WS /v1/realtime.

Duet

Two live sessions cross-piped into one conversation. This is how a persona gets tested against a difficult counterpart before a customer meets it.

​
The rest of the stack

Everything a conversation needs around it, on the same key and the same account.

Text to speech

Synthesis in a built-in or cloned voice. POST /v1/audio/speech.

Speech to text

Recorded audio to text, with timings. POST /v1/audio/transcriptions.

Voices

Clone from a few seconds of speech, or use the built-in library. POST /v1/voices.

Live translation

Interpretation as its own session type. You speak; the room hears you a beat later in another language.

​
Start

Quickstart

A key, one synthesized line, and a live conversation. Ten minutes.

What EESI does today

The honest map: every capability, the model behind it, and where it stops.

Models

The catalog, what each id serves, and why you pin the version.

Authentication

Keys, the two header forms, and browser tickets for a WebSocket.

​
Two things to know before you build

They explain most of the API surface, and getting either wrong looks like a bad model rather than a mistake.
The expensive mistake is assembling a conversation out of the file endpoints: slice the microphone, POST each slice to transcription, send the text to a chat model, synthesize the reply. That costs three round trips a turn, and it destroys turn boundaries — fixed slices cut words in half and tell the model nothing about whether a person has finished speaking.Audio you already have is a file. Audio arriving as someone speaks is a conversation. See Realtime interaction.
Cloned or built-in, a voice is always a piece of reference audio. Without one, the synthesis model samples a new voice per generation, and a call changes speaker between sentences. This is why voices are a resource you create rather than a string you pass. See Voices.
Everything here is reachable three ways: in the console, over the REST API, and through the SDKs. The speech endpoints also speak the OpenAI wire protocol, so an existing client moves with a base URL and a key — see OpenAI compatibility.