Speech API

The EESI Speech API is a drop-in replacement for OpenAI’s audio APIs. Point any OpenAI SDK at your EESI base URL with an sk-eesi- key and audio.speech, audio.transcriptions, models.list, and the Realtime WebSocket all work unchanged — backed by EESI’s own models (Supertonic, Kokoro, Parakeet, Whisper, OmniVoice, and a realtime speech-to-speech pipeline).

from openai import OpenAI

client = OpenAI(base_url="https://api.eesi.ai/v1", api_key="sk-eesi-...")

speech = client.audio.speech.create(model="supertonic", voice="F1", input="Hello from EESI.")
speech.stream_to_file("hello.mp3")

The only change from OpenAI is the base_url and the key. Everything below — fields, response formats, streaming events, and the error envelope ({"error": {"message", "type", "param", "code"}}) — matches OpenAI exactly.

Endpoints

Method	Endpoint	What it does
`POST`	`/v1/audio/speech`	Text → speech (streams audio bytes or SSE)
`POST`	`/v1/audio/transcriptions`	Speech → text (JSON / text / verbose / SSE stream)
`GET`	`/v1/models`	List available models (OpenAI shape)
`GET`	`/v1/audio/models`	List speech models with tier / streaming / cloning flags
`WS`	`/v1/realtime`	OpenAI Realtime protocol — live speech-to-speech
`POST`	`/v1/voices`	Clone a voice from a short reference clip
`GET`	`/v1/voices`	List your organization’s cloned voices
`DELETE`	`/v1/voices/{voice_id}`	Delete a cloned voice

Models

Model availability depends on what your deployment has configured — call GET /v1/models (or GET /v1/audio/models for capability flags) to see the live set. OpenAI model names are accepted as aliases so existing code keeps working.

Model	Kind	Tier	OpenAI aliases
`kokoro`	TTS	free	`tts-1`, `tts-1-hd`, `gpt-4o-mini-tts`
`supertonic`	TTS	free	—
`omnivoice`	TTS (voice cloning)	paid	—
`whisper-small`	STT	free	`whisper-1`, `gpt-4o-transcribe`
`parakeet-tdt-0.6b-v3`	STT (streaming)	free	—
`whisper-large-v3`	STT	paid	—
`eesi-s2s-mini`	Realtime	free	`gpt-realtime`, `gpt-4o-realtime-preview`
`eesi-s2s`	Realtime (voice cloning)	paid	—

OpenAI voice names (alloy, nova, …) map to the model’s default voice, so voice="alloy" never fails on a model that has its own preset voices.

Text to speech

POST /v1/audio/speech — full reference.

client.audio.speech.create(
    model="supertonic",       # or kokoro, omnivoice, or an OpenAI alias
    voice="F1",               # a preset, an OpenAI voice name, or a cloned ev_* id
    input="The quick brown fox.",
    response_format="mp3",    # mp3 | wav | pcm  (opus/aac/flac depend on the model)
    speed=1.0,                # 0.25–4.0
)

Set stream_format="sse" to receive audio as speech.audio.delta events instead of a raw byte stream.

curl https://api.eesi.ai/v1/audio/speech \
  -H "Authorization: Bearer sk-eesi-..." \
  -H "Content-Type: application/json" \
  -d '{"model":"supertonic","voice":"F1","input":"Hello from EESI.","response_format":"mp3"}' \
  --output hello.mp3

import OpenAI from "openai";
const client = new OpenAI({ baseURL: "https://api.eesi.ai/v1", apiKey: "sk-eesi-..." });

const res = await client.audio.speech.create({
  model: "supertonic", voice: "F1", input: "Hello from EESI.",
});

Transcription

POST /v1/audio/transcriptions — full reference. Multipart upload, up to 25 MB, with the exact OpenAI field set (file, model, language, prompt, response_format, temperature, timestamp_granularities[], stream).

with open("call.mp3", "rb") as f:
    result = client.audio.transcriptions.create(model="whisper-1", file=f)
print(result.text)

stream=true returns Server-Sent Events (transcript.text.delta / transcript.text.done). Models whose upstream can’t stream still honor stream=true — EESI runs the batch call and emits the transcript as a single delta, so streaming client code works against every model.

Realtime speech-to-speech

WS /v1/realtime?model=eesi-s2s-mini speaks the OpenAI Realtime protocol: send input_audio_buffer.append with base64 PCM16 @ 16 kHz, receive conversation.item.input_audio_transcription.delta, response.output_audio.delta, and barge-in via response.done (status: "cancelled", reason: "turn_detected"). Authentication accepts every channel OpenAI clients use:

Server SDKs send Authorization: Bearer sk-eesi-... on the handshake.
Browser clients send ?token=<clerk-jwt> / ?api_key=sk-eesi-..., or the openai-insecure-api-key.<key> WebSocket subprotocol.

async with client.realtime.connect(model="eesi-s2s-mini") as conn:
    await conn.session.update(session={"modalities": ["audio", "text"]})
    ...

Voice cloning

Clone a voice once and use it everywhere — the Playground, your agents, and the Realtime API. Upload 3–10 seconds of reference audio; EESI transcribes it automatically if you don’t supply a transcript.

curl https://api.eesi.ai/v1/voices \
  -H "Authorization: Bearer sk-eesi-..." \
  -F "file=@reference.wav" \
  -F "name=Sarah — Sales"
# → { "voice_id": "ev_1a2b3c4d", "status": "processing", ... }

Once status is ready, pass the ev_* id as the voice on any cloning-capable model:

client.audio.speech.create(model="omnivoice", voice="ev_1a2b3c4d", input="...")

See Create a voice for the full flow.

Billing

Free-tier models (kokoro, supertonic, whisper-small, parakeet-tdt-0.6b-v3, eesi-s2s-mini) are never metered. Paid-tier models draw from your prepaid credit balance and require credits to start; an empty balance returns 429 with code: "insufficient_quota", exactly like OpenAI. Self-hosted deployments with no billing configured never gate.

Resources

Authentication & Errors

Endpoints

Models

Text to speech

Transcription

Realtime speech-to-speech

Voice cloning

Billing

​Endpoints

​Models

​Text to speech

​Transcription

​Realtime speech-to-speech

​Voice cloning

​Billing

Endpoints

Models

Text to speech

Transcription

Realtime speech-to-speech

Voice cloning

Billing