OpenAI-Compatible Text to Speech API

Generate speech from text through Eden AI’s OpenAI-compatible speech endpoint. Point any OpenAI client at https://api.edenai.run/v3 and client.audio.speech works as a drop-in replacement for OpenAI’s text to speech.

Overview

Method	Path	Purpose
`POST`	`/v3/audio/speech`	Synthesize text into audio. Returns the raw audio bytes
`GET`	`/v3/audio/speech/models`	Speech models available in your region, as `provider/model` ids
`GET`	`/v3/audio/voices`	Voices the speech models accept, with language and gender

The response body is the audio file itself, as with OpenAI. Eden’s cost (USD) and provider come back in the x-edenai-cost and x-edenai-provider response headers.

Speech models are only available on /v3/audio/speech. They are not in GET /v3/models, which lists chat models.

This is different from Expert Model text to speech, which runs through /v3/universal-ai with each provider’s native options. Use /v3/audio/speech when you want an OpenAI-compatible, drop-in speech surface.

Models

Provider	Model string
OpenAI	`openai/tts-1`, `openai/tts-1-hd`
ElevenLabs	`elevenlabs/eleven_multilingual_v2`, `elevenlabs/eleven_v3`
Amazon Polly	`amazon/standard`, `amazon/neural`
Google	`vertex/chirp`

Call GET /v3/audio/speech/models for the full list available in your region, with the per-character price of each model.

amazon/generative and amazon/long-form are in the model list, but no voice in GET /v3/audio/voices is available for them yet, so a request to either returns 400 when it names no voice or a voice from that list.

Generating speech

import requests

url = "https://api.edenai.run/v3/audio/speech"
headers = {
    "Authorization": "Bearer YOUR_API_KEY",
    "Content-Type": "application/json",
}

payload = {
    "model": "openai/tts-1",
    "input": "Hello! This sentence was spoken through Eden AI.",
    "voice": "alloy",
}

response = requests.post(url, headers=headers, json=payload)
response.raise_for_status()

with open("speech.mp3", "wb") as f:
    f.write(response.content)

print("Provider:", response.headers["x-edenai-provider"])
print("Cost (USD):", response.headers.get("x-edenai-cost"))

curl -X POST https://api.edenai.run/v3/audio/speech \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/tts-1",
    "input": "Hello! This sentence was spoken through Eden AI.",
    "voice": "alloy"
  }' \
  --fail-with-body \
  --output speech.mp3

With --fail-with-body, an error makes curl exit with a non-zero status. The JSON error is still written to speech.mp3, so read that file to see why the request failed. The official OpenAI SDK works unchanged once base_url points at Eden AI:

from openai import OpenAI

client = OpenAI(base_url="https://api.edenai.run/v3", api_key="YOUR_API_KEY")

with client.audio.speech.with_streaming_response.create(
    model="openai/tts-1",
    voice="alloy",
    input="Hello! This sentence was spoken through Eden AI.",
) as response:
    response.stream_to_file("speech.mp3")

To read the cost and provider headers from the SDK, use with_raw_response:

from openai import OpenAI

client = OpenAI(base_url="https://api.edenai.run/v3", api_key="YOUR_API_KEY")

raw = client.audio.speech.with_raw_response.create(
    model="elevenlabs/eleven_multilingual_v2",
    voice="kwhMCf63M8O3rCfnQ3oQ",  # Caroline, a French voice
    input="Bonjour ! Cette phrase a été prononcée via Eden AI.",
)
raw.parse().write_to_file("bonjour.mp3")

print("Provider:", raw.headers["x-edenai-provider"])
print("Cost (USD):", raw.headers.get("x-edenai-cost"))

Request parameters

Field	Required	Description
`model`	yes	`provider/model`, e.g. `openai/tts-1`. The model name alone (`tts-1`) also works, through Provider Routing
`input`	yes	The text to synthesize
`voice`	no	A voice `id` or `name` from `GET /v3/audio/voices`. Omit it to let Eden AI pick one. The OpenAI SDK requires it, so only a raw HTTP request can leave it out
`response_format`	no	`mp3`, `opus`, `aac`, `flac`, `wav` or `pcm`. Not every provider returns every format: see Output formats
`speed`	no	Playback speed. OpenAI models accept 0.25 to 4.0; other providers may use a different range or ignore it
`instructions`	no	Guidance on voice and delivery style. None of the models listed above supports it: it does not work with OpenAI’s `tts-1` and `tts-1-hd`, and the other providers drop it

Unknown fields are ignored.

Output formats

The Content-Type of the response follows the response_format you request, not the audio the provider actually returns. Every provider returns mp3; the other formats depend on the provider:

`response_format`	`Content-Type`	OpenAI	ElevenLabs	Amazon Polly	Google
`mp3`	`audio/mpeg`	MP3	MP3	MP3	MP3
`opus`	`audio/ogg`	Opus	Opus	Ogg Vorbis	Opus
`aac`	`audio/aac`	AAC	Error	MP3	MP3
`flac`	`audio/flac`	FLAC	Error	MP3	FLAC
`wav`	`audio/wav`	WAV	Error	PCM, no WAV header	WAV
`pcm`	`audio/pcm`	PCM	PCM	PCM	WAV
omitted	`audio/mpeg`	MP3	MP3	MP3	WAV

With vertex/chirp, set response_format to mp3 if you need MP3: when it is omitted, Google returns WAV under an audio/mpeg header.

Choosing a voice

GET /v3/audio/voices lists the voices the speech models accept. Each voice names the models that take it:

import requests

url = "https://api.edenai.run/v3/audio/voices"
headers = {"Authorization": "Bearer YOUR_API_KEY"}
params = {"model": "elevenlabs/eleven_multilingual_v2", "gender": "female"}

response = requests.get(url, headers=headers, params=params)
voices = response.json()["data"]

for voice in voices[:5]:
    print(voice["id"], "|", voice["name"], "|", voice["gender"])

curl "https://api.edenai.run/v3/audio/voices?model=openai/tts-1" \
  -H "Authorization: Bearer YOUR_API_KEY"

{
  "object": "list",
  "data": [
    {
      "id": "alloy",
      "object": "voice",
      "name": "alloy",
      "provider": "openai",
      "models": ["openai/tts-1"],
      "languages": [],
      "gender": "unknown",
      "preview_url": null,
      "default": true
    }
  ]
}

Query parameter	Description
`model`	Only the voices this model accepts, e.g. `openai/tts-1`. Each voice’s `models` then lists only that model. Without it, the voices of every speech model available to you
`language`	Only the voices speaking this language. A bare code (`fr`) matches every region; `fr-CA` matches that locale only. Multilingual voices are always kept
`gender`	`male`, `female` or `unknown`

How the voice field of a speech request is matched:

An id or a full name from the list is matched without regard to case: "alloy" and "Alloy" are the same OpenAI voice. Provider ids are shorter and stable, so prefer them for voices with long display names, such as most ElevenLabs voices.
When voice is omitted, the model’s default voice ("default": true) is used. A model with no default voice, such as vertex/chirp, gets the first voice it accepts in alphabetical order, which may not speak the language of your text, so name a voice for it.
A voice of the model’s provider that the model does not accept returns 400, with the voices the model does accept in the error message.
Any other voice, including one that belongs to another provider, is sent to the provider unchanged. Cloned or account-specific voice ids still reach the provider this way. Amazon models read OpenAI’s voice names as Polly voices, so alloy is spoken by Joanna.

Listing speech models

cURL

curl https://api.edenai.run/v3/audio/speech/models \
  -H "Authorization: Bearer YOUR_API_KEY"

Each entry carries its pricing, e.g. {"input_cost_per_character": 0.000015} for openai/tts-1. The endpoint takes the same view parameter as the other model listings.

Pricing

Speech is billed on the input text, per character (or per token for models priced that way), at the rate GET /v3/audio/speech/models shows. The amount charged for a request is in its x-edenai-cost header.

Limitations

No @edenai routing. "model": "@edenai" returns 400; name a model.
No fallbacks. The field is ignored, like any unknown field: only model is tried, and a failure is not retried on another model.
No streaming playback. The audio is returned once the provider has synthesized all of it. The SDK’s with_streaming_response writes it to disk without holding it in memory, but does not start any sooner.
No edenai_metadata block. The body is raw audio, so the request metadata block is not returned, even with x-edenai-metadata: enabled; read the x-edenai-provider and x-edenai-cost headers instead.

V3 Documentation

Quick Start

Overview

LLMs

Expert Models

General

Organization

Data Governance

Integrations

OpenAI-Compatible Text to Speech API

Overview

Models

Generating speech

Request parameters

Output formats

Choosing a voice

Listing speech models

Pricing

Limitations

Next Steps

Speech to Text

Expert Text to Speech

​Overview

​Models

​Generating speech

​Request parameters

​Output formats

​Choosing a voice

​Listing speech models

​Pricing

​Limitations

​Next Steps

Speech to Text

Expert Text to Speech

Overview

Models

Generating speech

Request parameters

Output formats

Choosing a voice

Listing speech models

Pricing

Limitations

Next Steps