OmniVoice

Name in the API

OmniVoice

Base model

k2-fsa/OmniVoice

Type

Text to speech

Input → output

text → audio

Availability

Dedicated

Hardware

A 12 GB slice of a GPU per instance

Voices

alloy, echo, fable, nova, onyx, shimmer: fixed voices, cloned from short reference recordings.

default: no reference voice. The model picks a voice itself, so it can sound different in every request.

voice is required. An unknown name is rejected with error 400, and the message lists the available voices.

Language

language with the OmniVoice language ID, e.g. "de" or "en". Optional.

Seed

seed (an integer from 0 to 2^63 − 1): the same seed, text and voice give exactly the same audio on the same instance. Our instances run on different GPU types, so two requests can still sound slightly different.

Long texts

Texts longer than about 30 seconds of speech are split at sentence boundaries into parts of about 15 seconds, which are joined again. Later parts keep the voice of the first part. We tested about 400 words (2 minutes of audio).

Change the sizes per request with extra_params: {"audio_chunk_duration": 15, "audio_chunk_threshold": 30} (seconds).

Endpoint

/v1/audio/speech

Usage

from openai import OpenAI

client = OpenAI(base_url="https://llm.scads.ai/v1", api_key="<your API key>")

audio = client.audio.speech.create(
    model="OmniVoice",
    voice="nova",
    input="Guten Tag aus Dresden!",
    response_format="wav",
    extra_body={"language": "de", "seed": 42},
)
with open("hello.wav", "wb") as f:
    f.write(audio.content)