Migrating to EESI

The speech endpoints implement the OpenAI wire protocol, so migrating an existing integration is a base URL and a key. The official OpenAI SDKs work unchanged; nothing else in your code has to move.

The two lines

from openai import OpenAI

client = OpenAI(
-   api_key=os.environ["OPENAI_API_KEY"],
+   api_key=os.environ["EESI_API_KEY"],
+   base_url="https://api.eesi.ai/v1",
)

const client = new OpenAI({
- apiKey: process.env.OPENAI_API_KEY,
+ apiKey: process.env.EESI_API_KEY,
+ baseURL: "https://api.eesi.ai/v1",
});

Get a key from API keys — see Get your API key.

What maps to what

You were calling	Call this	Notes
`audio.speech.create`	Same	`tts-1` / `gpt-4o-mini-tts` → an EESI `tts` model id
`audio.transcriptions.create`	Same	`whisper-1` resolves to `nur-stt-v1`
`chat.completions.create`	Same	Text LLM, unchanged
Realtime over WebSocket	`wss://api.eesi.ai/v1/realtime`	Speech-to-speech

Don’t hardcode a model id. Speech tiers are opt-in per environment: a backend with no configured URL is hidden entirely, so an id that works in one deployment 404s in another. Read GET /v1/audio/models and pick from what it returns — that list is the contract.

curl https://api.eesi.ai/v1/audio/models \
  -H "Authorization: Bearer $EESI_API_KEY"

Voices are not names

This is the one real difference in the request body. OpenAI ships a fixed set of voice names (alloy, nova, …). On EESI a voice is always a stored reference clip — built-in voices carry human names, and cloned voices are ev_-prefixed ids from your own library:

{ "model": "nur-tts-v1", "voice": "Nur", "input": "..." }
{ "model": "nur-tts-v1", "voice": "ev_a68b5c50", "input": "..." }

That invariant is the reason a long call keeps one voice: without reference audio the model samples a new voice on every generation and a conversation drifts between speakers. See Voices.

What you gain

These are extra fields on the same endpoint — ignored by models that don’t support them, so adding one never breaks a fallback path.

Field	Does
`instructions`	Designs a voice from a phrase instead of naming one
`language`	Sets the synthesis language explicitly
`normalize_text`	Reads numbers, dates and currency as words
`speed`	`0.25`–`4.0`
`source`	Labels the request in your generation history

Two response headers have no OpenAI equivalent:

X-Synthetic-Audio — every response declares itself AI-generated, which is an EU AI Act Article 50(2) obligation rather than a nicety.
X-EESI-Generation-Id — the history entry this audio became.

You also get voice cloning (/v1/voices), diarized transcripts, and a replayable history of every generation, none of which have an OpenAI counterpart.

What to check before you cut over

Formats. Each model advertises its own formats list. mp3 is universal; opus, aac, flac, wav and pcm depend on the model.
Billing. EESI meters prepaid credits, and requests fail with 402 at zero rather than invoicing later. See Billing.
Concurrency. Realtime sessions are capped per organization by plan, and a refused session returns cleanly rather than degrading a live one.
Rate limits. 429 means the plan’s request rate; retry with backoff.

Where the compute runs

The API sits in front of EESI’s own GPUs — see Where the compute runs.

Get started

Speech

Realtime

Telephony

Care

Platform

Migrating to EESI

The two lines

What maps to what

Voices are not names

What you gain

What to check before you cut over

Where the compute runs

​The two lines

​What maps to what

​Voices are not names

​What you gain

​What to check before you cut over

​Where the compute runs

The two lines

What maps to what

Voices are not names

What you gain

What to check before you cut over

Where the compute runs