OmniVoice

POST /v1/audio/speech model: OmniVoice

This model is not on /v1/chat/completions

OmniVoice answers on POST /v1/audio/speech and returns audio bytes, not JSON. messages, tools, max_tokens and stream do not apply.

Non-commercial use only

The OmniVoice model weights are licensed CC-BY-NC, so the audio may not be used commercially. Its audio tokenizer is Boson AI’s Higgs Audio 2 (see License).

At a glance

Endpoint	POST /v1/audio/speech
Input	text, 1–4,096 characters
Output	one audio file, 24 kHz mono: wav (default), mp3, opus, flac or pcm
Streaming	❌ — the whole file arrives in one response
Voice	a preset, a free-text description, or a cloned reference clip
Billing	per input character; the input rate on the card applies per million characters
Stability	experimental

Request

Field
model	required	OmniVoice
input	required	the text to speak, 1–4,096 characters
voice	optional	a preset, see below. Default auto
instructions	optional	describe the voice in words, up to 512 characters, e.g. "male, low pitch, british accent". Takes precedence over voice
response_format	optional	wav (default), mp3, opus, flac, pcm
speed	optional	0.25–4.0
language	optional	a language code such as en or zh; omit it to let the model decide from the text
ref_audio	optional	a voice to clone: base64 or a data: URL of a WAV, FLAC, OGG or MP3 clip, at most 30 s (3–10 s works best)
ref_text	optional	the transcript of ref_audio
duration	optional	target length of the output in seconds, up to 600
num_step	optional	diffusion steps, 1–64. Fewer is faster and rougher
seed	optional	the same seed with the same request returns identical audio

Presets for voice: auto, female, male, female-low, male-low, female-british, male-british, female-american, male-american. OpenAI voice names such as alloy are rejected with unsupported_voice.

A preset fixes a style, not a speaker

voice and instructions describe what the voice should sound like, so two requests with the same preset can come back as two different-sounding people. For one consistent voice across requests, send the same ref_audio, or pin seed.

Cloning a voice

Send a short clip in ref_audio and leave out voice and instructions — combining them is a 400 (invalid_parameters). Include ref_text: without it the clip is transcribed first, which in our tests took about 17 s instead of 4 s. The whole request body must stay under 5 MB.

Only clone a voice you have the right to use.

Response

The body is the audio file itself. Content-Type follows response_format:

response_format	Content-Type
wav	audio/wav	16-bit PCM
mp3	audio/mpeg
opus	audio/ogg	Opus in an Ogg container
flac	audio/flac
pcm	audio/pcm	raw 16-bit little-endian samples, 24 kHz mono, no header

Errors are JSON in the usual shape, with these codes:

Status	code	Cause
400	invalid_parameters	input empty or over 4,096 characters, ref_audio combined with voice/instructions, or ref_text without ref_audio
400	unsupported_voice	voice is not one of the presets
400	unsupported_response_format	e.g. aac
400	invalid_ref_audio	the clip cannot be decoded, is empty, or is longer than 30 s
400	unsupported_stream_format	stream_format: "sse" was sent
503	server_busy	too many requests queued; retry after a short wait

Example

curl https://developer.amd.com.cn/radeon/api/v1/audio/speech \
  -H "Authorization: Bearer $RADEON_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "OmniVoice",
    "input": "Hello from Radeon Cloud.",
    "voice": "female",
    "response_format": "mp3"
  }' \
  -o speech.mp3

Cloning a voice from sample.wav:

import base64, requests

ref = base64.b64encode(open("sample.wav", "rb").read()).decode()
r = requests.post(
    "https://developer.amd.com.cn/radeon/api/v1/audio/speech",
    headers={"Authorization": f"Bearer {API_KEY}"},
    json={
        "model": "OmniVoice",
        "input": "This sentence is spoken in the voice from the clip.",
        "ref_audio": f"data:audio/wav;base64,{ref}",
        "ref_text": "The words spoken in sample.wav.",
        "response_format": "wav",
    },
    timeout=120,
)
r.raise_for_status()
open("cloned.wav", "wb").write(r.content)

One file per request, generated in full first

Nothing is returned until the whole clip has been generated — about 1–4 s for a sentence. Requests are served one at a time and queue behind each other, so long texts and bursts wait longer.

License

The OmniVoice code is Apache-2.0 and the model weights are CC-BY-NC (non-commercial). The audio tokenizer is licensed under the Boson Higgs Audio 2 Community License, which asks us to show:

Built with Higgs Materials licensed from Boson AI USA, Inc., Copyright Boson AI USA, Inc., All Rights Reserved and Meta Llama 3 licensed under the Meta Llama 3 Community License, Copyright Meta Platforms, Inc., All Right Reserved. Based on Meta Llama 3.