Create Speech

POST /v1/audio/speech Models: Speech View .md

Generate speech from text. OpenAI's request shape plus OpenRouter's, so both SDKs work unchanged. The response body is the audio itself.

Voices: voice selects the mode as well as the voice. A built-in speaker (GET /v1/audio/voices?model= lists them) speaks the input, and instructions can direct its style. design creates a new voice from the description in instructions. clone (or leaving voice out) with input_references speaks in the voice of a reference clip: send it base64-encoded in input_audio.data, with its transcript in a text part for the closest match. Clip URLs are not accepted. A cloned voice cannot follow instructions.

Formats: response_format is mp3 (the default), opus, flac, wav or pcm (24 kHz 16-bit mono, labelled audio/pcm;rate=24000;channels=1). aac returns 400.

Streaming: stream_format audio (the default) returns the audio; pcm at speed 1 streams as it is generated. sse returns speech.audio.delta events with base64 audio and a final speech.audio.done event with usage; pcm at speed 1 streams, other formats arrive as one delta. A failure mid-stream ends an SSE stream with a speech.audio.error event.

Billing: Per second of audio generated, before speed, rounded up. The seconds and cost are in the X-Usage-Seconds and X-Usage-Cost headers, or in the speech.audio.done usage; a raw pcm stream reports neither.

OpenRouter's provider, user, session_id and trace are accepted and ignored, except that provider.options under our OpenRouter slug can carry instructions, language, speed, seed, temperature, top_p and top_k.

Body

application/json

Text-to-Speech Request Payload

model string required

The text-to-speech model, e.g. qwen3-tts-12hz-1.7b, chatterbox or auk.
input string required

The text to speak, up to the model's max_prompt_length in characters.
voice string or object

One of the model's voices (GET /v1/audio/voices), as a string or as OpenAI's {"id": ...}: for Qwen3-TTS a built-in speaker, design or clone; for Chatterbox default or clone; for AuK default, design or clone. Defaults to the model's default voice, or clone when input_references is sent.
One of 2
- string
- object
  - id string required
instructions string

Style for a built-in speaker, or the description of the voice for design. Not accepted with clone, nor by a model that does not follow instructions (Chatterbox).
input_references array of objects

OpenRouter's reference parts for clone. One input_audio clip and, optionally, its transcript as a text part (Chatterbox clones from the audio alone and does not use it).
3 fields
- type string required
  
  input_audiotext
- input_audio object
  2 fields
  - data string required
    
    The clip, base64-encoded or a base64 data URI. Up to 15 MB decoded. URLs are not accepted.
  - format string
    
    wavmp3flacoggm4awebm
- text string
  
  What the clip says.
  
  at most 10000 characters
language string

The language of the input, as auto (the default), an ISO 639-1 code, a BCP 47 tag or a name. For chatterbox it picks the checkpoint: auto or English is the English Original, any other of its 23 languages is Multilingual V3. Nothing is guessed from the text.
response_format string

default mp3mp3opusflacwavpcm
stream_format string

default audioaudiosse
speed number

default 1range 0.25–4
seed integer
temperature number

range 0–2
top_p number

The engine takes values above 0, so 0 is sent as 0.000001, which keeps only the likeliest token.

range 0–1
top_k integer

at least -1
exaggeration number

Emotion intensity, for a model whose catalog entry lists it (Chatterbox, default 0.5). Higher is more expressive and tends to speak faster.

range 0.25–2
cfg_weight number

Guidance toward the voice, for a model whose catalog entry lists it (Chatterbox, default 0.5). Lower slows the pacing; 0 keeps a voice's accent out of another language, and is the default for Chatterbox's built-in voice outside English.

range 0–1
num_inference_steps integer

Denoising steps, for a model whose catalog entry lists it (AuK, default 32). Fewer is faster and coarser.

range 4–64
guidance_scale number

How closely the speech follows its prompt and reference voice, for a model whose catalog entry lists it (AuK, default 2).

range 1–5

Responses

The audio, or SSE events when stream_format is sse.

audio/mpeg, audio/ogg, audio/flac, audio/wav, audio/pcm, text/event-stream

curl https://api.vultrinference.com/v1/audio/speech \
  -H "Authorization: Bearer $VULTR_INFERENCE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3-tts-12hz-1.7b",
    "input": "Welcome to Vultr Serverless Inference.",
    "voice": "aiden",
    "instructions": "Warm and upbeat, like a radio host."
  }'

import os
import requests

response = requests.post(
    "https://api.vultrinference.com/v1/audio/speech",
    headers={"Authorization": f"Bearer {os.environ['VULTR_INFERENCE_API_KEY']}"},
    json={
        "model": "qwen3-tts-12hz-1.7b",
        "input": "Welcome to Vultr Serverless Inference.",
        "voice": "aiden",
        "instructions": "Warm and upbeat, like a radio host.",
    },
)
print(response.status_code)

const response = await fetch("https://api.vultrinference.com/v1/audio/speech", {
  method: "POST",
  headers: {
    Authorization: `Bearer ${process.env.VULTR_INFERENCE_API_KEY}`,
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    "model": "qwen3-tts-12hz-1.7b",
    "input": "Welcome to Vultr Serverless Inference.",
    "voice": "aiden",
    "instructions": "Warm and upbeat, like a radio host."
  }),
});
console.log(response.status);