Venice Music / Async Audio

Music, sound effects and character-priced voice generation are asynchronous:

Method	Path	Auth	Notes
POST	/api/v1/audio/quote	None required (key optional)	Price in USD.
POST	/api/v1/audio/queue	Bearer key or x402 (SIWX)	Charges and enqueues → queue_id. 40 req/min per user.
POST	/api/v1/audio/retrieve	Bearer key or x402 (SIWX)	Status JSON or the audio bytes. 120 req/min per user.
POST	/api/v1/audio/complete	Bearer key or x402 (SIWX)	Delete the stored media.

For short synchronous text-to-speech use venice-audio-speech. Voice-changer (speech-to-speech) models are refused on these four endpoints with a 400 pointing at /audio/voice-changer/* (callers who can't see the model get 404 instead) — see venice-audio-voice-changer.

Use when

• You need songs, jingles, score, soundscapes, sound effects, or long narration.
• The model uses duration-, per-second-, per-job- or character-based pricing and you want a price before submitting.
• Generation takes long enough that a synchronous call would time out.

Models

Query GET /models?type=music for the current list and each model's model_spec. Representative ids (all in the live list):

Kind	Examples
Instrumental / songs	elevenlabs-music, elevenlabs-music-v2-5, lyria-3-pro, sonilo-v1-1-music, stable-audio-25
Songs with lyrics	minimax-music-v25, minimax-music-v26, minimax-music-v2 (lyrics required), ace-step-15 (lyrics optional)
Sound effects	elevenlabs-sound-effects-v2 (supports loop), sonilo-v1-1-sound-effects, mmaudio-v2-text-to-audio
Voice (text in prompt)	elevenlabs-tts-v4, elevenlabs-tts-v4-turbo, elevenlabs-tts-v3, elevenlabs-tts-multilingual-v2, seed-audio-1-0

Lifecycle

1. POST /audio/quote — price it first

curl https://api.venice.ai/api/v1/audio/quote \
  -H "Content-Type: application/json" \
  -d '{ "model": "elevenlabs-music", "duration_seconds": 60 }'

Response: {"quote": 0.69} (USD). No API key needed; sending one lets you price models only your account can see.

Field	Notes
model	Required.
duration_seconds	Integer or numeric string. Only for models that expose duration metadata (min_duration / max_duration / duration_options) — rejected otherwise. Omit to price the model's default_duration.
character_count	Integer ≤ prompt_character_limit. Required for models priced by per_thousand_characters.

Unknown fields → 400.

2. POST /audio/queue — enqueue

curl https://api.venice.ai/api/v1/audio/queue \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "elevenlabs-music",
    "prompt": "Uplifting indie-folk acoustic track, 120 BPM, major key.",
    "duration_seconds": 60,
    "force_instrumental": true
  }'

Song with lyrics (minimax-music-v25 supports lyrics, the optimizer and instrumental mode, but no duration_seconds):

{
  "model": "minimax-music-v25",
  "prompt": "Warm indie-pop ballad, female vocals, acoustic guitar.",
  "lyrics_prompt": "[Verse]\nWalking through the city lights...\n[Chorus]\nWe are the dreamers..."
}

Response: { "model": "...", "queue_id": "...", "status": "QUEUED" }.

The body is strict: every optional field below is rejected with 400 when the model's model_spec says it isn't supported.

Field	Notes
model	Required.
prompt	Required. Between min_prompt_length (default 10) and prompt_character_limit; trimmed. For the voice models this is the text to speak.
lyrics_prompt	Up to lyrics_character_limit (default 4096). Required when lyrics_required=true; rejected when supports_lyrics=false.
duration_seconds	Integer or numeric string. Must be one of duration_options when present, else within min_duration–max_duration. Defaults to default_duration.
force_instrumental	supports_force_instrumental=true only.
lyrics_optimizer	Auto-writes lyrics from prompt. supports_lyrics_optimizer=true only; lyrics_prompt must then be empty.
loop	Render a seamless loop (end splices into start). supports_loop=true only — currently elevenlabs-sound-effects-v2.
voice	Voice-enabled models only. One of voices; defaults to default_voice. Models with supports_custom_voice_id=true (the ElevenLabs TTS models) also accept a raw ElevenLabs Voice ID.
language_code	ISO 639-1. supports_language_code=true only — no model in the current list sets it.
speed	supports_speed=true only, within min_speed–max_speed.

Model-specific rules also apply: minimax-music-v25 needs a lyrics_prompt of at least 10 chars unless force_instrumental or lyrics_optimizer is true; minimax-music-v26 needs the same unless force_instrumental is true; minimax-music-v2 needs a non-blank lyrics_prompt.

3. POST /audio/retrieve — poll / download

curl https://api.venice.ai/api/v1/audio/retrieve \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"elevenlabs-music","queue_id":"..."}' \
  --output track.mp3

• Still running: 200 JSON {"status":"PROCESSING","average_execution_time":<ms, P80 estimate>,"execution_duration":<ms since queued>}.
• Done: 200 with the audio bytes. Content-Type is the audio type; headers x-venice-audio-format, x-venice-inference-time (s), x-venice-model-id, x-venice-model-name, and for Seed Audio also x-venice-audio-duration and x-venice-audio-subtitle.
• delete_media_on_completion: true deletes the media after this download, so you can skip step 4.

If generation fails (content policy, capacity, provider validation), the charge is refunded (except a DIEM charge from a previous epoch) and the error is returned here.

4. POST /audio/complete — cleanup

curl https://api.venice.ai/api/v1/audio/complete \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"elevenlabs-music","queue_id":"..."}'

Returns {"success": true} once the stored media is deleted (false if the delete didn't go through). Use it after you've saved the bytes, unless you retrieved with delete_media_on_completion: true.

Full loop (TypeScript)

import fs from 'node:fs/promises'

const base = 'https://api.venice.ai/api/v1'
const headers = {
  Authorization: `Bearer ${process.env.VENICE_API_KEY}`,
  'Content-Type': 'application/json',
}

async function generateTrack() {
  // 1. Quote
  const quote = await fetch(`${base}/audio/quote`, {
    method: 'POST', headers,
    body: JSON.stringify({ model: 'elevenlabs-music', duration_seconds: 60 }),
  }).then(r => r.json())
  console.log('price:', quote.quote)

  // 2. Queue
  const { queue_id, model } = await fetch(`${base}/audio/queue`, {
    method: 'POST', headers,
    body: JSON.stringify({
      model: 'elevenlabs-music',
      prompt: 'Uplifting indie-folk acoustic track, 120 BPM.',
      duration_seconds: 60,
      force_instrumental: true,
    }),
  }).then(r => r.json())

  // 3. Poll
  while (true) {
    const res = await fetch(`${base}/audio/retrieve`, {
      method: 'POST', headers,
      body: JSON.stringify({ model, queue_id, delete_media_on_completion: true }),
    })
    if (!res.ok) throw new Error(`retrieve failed: ${res.status} ${await res.text()}`)
    const ct = res.headers.get('content-type') ?? ''
    if (!ct.startsWith('application/json')) {
      await fs.writeFile('track.mp3', Buffer.from(await res.arrayBuffer()))
      break
    }
    const { status } = await res.json()
    if (status !== 'PROCESSING') throw new Error(`unexpected ${status}`)
    await new Promise(r => setTimeout(r, 3000))
  }
  // delete_media_on_completion: true made /audio/complete unnecessary
}

Capability probing

Each GET /models?type=music entry's model_spec exposes:

• supports_lyrics, lyrics_required, lyrics_character_limit, supports_lyrics_optimizer
• supports_force_instrumental, supports_loop, supports_language_code
• supports_speed, default_speed, min_speed, max_speed
• voices[], default_voice, supports_custom_voice_id
• duration_options[], min_duration, max_duration, default_duration
• min_prompt_length, prompt_character_limit
• supported_formats, default_format (the output container — informational, not a request field)
• voice_changer: true marks speech-to-speech models that belong on /audio/voice-changer/*
• pricing, one of:
  • durations — { "<tier>": { usd, diem, min_seconds, max_seconds } } (e.g. elevenlabs-music, ace-step-15)
  • generation — flat per job (e.g. minimax-music-v25, lyria-3-pro, stable-audio-25)
  • per_second — per generated second (e.g. elevenlabs-sound-effects-v2, sonilo-v1-1-music, seed-audio-1-0)
  • per_thousand_characters — by prompt length (the ElevenLabs TTS models)

Errors

Code	Meaning
400	Schema error (strict body), unsupported option for the model, bad duration_seconds, lyrics_optimizer + lyrics_prompt, voice-changer model on these endpoints, a provider-side validation failure reported on retrieve (refunded), or an unknown / foreign queue_id on retrieve/complete ("Request ID is invalid."). Voice errors include details.supported_voices.
401	Authentication failed.
402	Insufficient balance. Bearer → {"error":"Insufficient USD or Diem balance…"}, or `"API key USD
403	A PRIVATE_ONLY key calling an anonymized model, or region restriction.
404	Unknown model; or on retrieve, media not found / expired / already deleted.
422	Content policy violation (queue or retrieve). May include suggested_prompt. Charge refunded (except a DIEM charge from a previous epoch).
429	Rate limited (40/min queue, 120/min retrieve, per user).
500	Inference failure.
503	Model at capacity — retry later.

See venice-errors for body shapes.

Gotchas

• Quote before queue. Queue charges up front (credits) or checks your x402 balance against the quote. With an API key, compare the quote to data.balances from GET /api_keys/rate_limits, which works with an INFERENCE key and is already capped at the key's spend limit; the request is charged to the first currency that covers the whole quote (DIEM, then earned credits, then bundled credits, then USD; balances doesn't list earned credits). With a wallet, use /x402/balance/....
• Sending an unsupported option (lyrics_prompt, voice, speed, language_code, loop, duration_seconds, …) is a 400, not a silent no-op. Build the body from model_spec.
• Store queue_id and model — every later call needs both.
• Media is ephemeral. Save the bytes on retrieve; after complete (or delete_media_on_completion) the audio is gone.
• seed-audio-1-0 takes no duration_seconds: it reserves its 120 s output cap at queue time and settles on the actual length when done.
• Poll every 2–5 s; use average_execution_time to pick the first delay. Faster polling doesn't speed the job up and eats the 120/min retrieve limit.