venice-audio-speech

Venice TTS (/audio/speech)

POST /api/v1/audio/speech converts text to an audio stream or file. OpenAI-compatible — the OpenAI SDK's audio.speech.create() works as a drop-in.

Use when

• You want narration, voice replies, or UI audio from text.
• You need a specific voice family (ElevenLabs, Kokoro, xAI, Qwen 3, Orpheus, Chatterbox, MiniMax, Inworld, Gemini Flash, Gradium).
• You want streaming audio returned as it is generated.
• You need style/emotion control on supported models, or synthesis in a cloned voice.

For music, sound effects and the character-priced ElevenLabs TTS models — v3, v4, v4 Turbo, Multilingual v2 — (async), see venice-audio-music. For transcription (audio → text), see venice-audio-transcription.

Minimal request

curl https://api.venice.ai/api/v1/audio/speech \
  -H "Authorization: Bearer $VENICE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{

Response is the raw audio. Content-Type is audio/mpeg, audio/opus, audio/aac, audio/flac, audio/wav or audio/pcm depending on the format actually produced.

Request schema

The body is strict — unknown fields return 400.

language by model: Qwen 3 → full names (English, Chinese, …; default auto); xAI → ISO 639-1 (en), passed through as given, so send a valid code (default auto); ElevenLabs Turbo → ISO 639-1 (values longer than 5 chars dropped, shorter ones forwarded); MiniMax → full names (sent as a language boost, unchecked); Gemini Flash → full locale strings such as English (US) or Japanese (Japan) (anything else dropped). Kokoro, Inworld, Chatterbox, Orpheus and Gradium ignore it.

Models

Every id below is in the live GET /models?type=tts list. Prices are model_spec.pricing.input.usd = USD per 1M input characters.

Always inspect GET /models?type=tts before calling: model_spec.voices (authoritative voice list), supported_formats, default_format, supports_custom_voice_id, privacy, pricing, and — on cloning models — voice_cloning. Per-model parameter support (prompt / temperature / top_p / language) is not exposed on /models; use the table above.

A key with modelPrivacy: PRIVATE_ONLY gets 403 on the anonymized models (PRIVATE_TEXT keys are not restricted here).

Voices (from model_spec.voices)

Case-sensitive. Omit voice to get the model's default.

• Kokoro — <lang><gender>_<name>: a American, b British, z Chinese, f French, h Hindi, i Italian, j Japanese, p Portuguese, e Spanish; f/m gender. Examples: af_sky, af_bella, af_heart, am_adam, am_michael, bf_emma, bm_george, ff_siwis, jf_alpha, zf_xiaoxiao, ef_dora, pm_alex (54 total).
• Qwen 3 — Vivian, Serena, Ono_Anna, Sohee, Uncle_Fu, Dylan, Eric, Ryan, Aiden
• xAI — eve, ara, rex, sal, leo, altair, atlas, carina, castor, celeste, cosmo, helios, helix, iris, kepler, lumen, luna, lux, naksh, orion, perseus, rigel, sirius, ursa, zagan, zenith
• Orpheus — tara, leah, jess, mia, zoe, leo, dan, zac
• Inworld — Craig, Ashley, Olivia, Sarah, Elizabeth, Priya, Alex, Edward, Theodore, Ronald, Mark, Hades, Luna, Pixie
• Chatterbox — Aurora, Britney, Siobhan, Vicky, Blade, Carl, Cliff, Richard, Rico
• ElevenLabs Turbo — Rachel, Aria, Sarah, Laura, Charlotte, Alice, Matilda, Jessica, Lily, Roger, Charlie, George, Callum, River, Liam, Will, Eric, Chris, Brian, Daniel, Bill — or any ElevenLabs Voice ID
• MiniMax — WiseWoman, FriendlyPerson, InspirationalGirl, CalmWoman, LivelyGirl, LovelyGirl, SweetGirl, ExuberantGirl, DeepVoiceMan, CasualGuy, PatientMan, YoungKnight, DeterminedMan, ImposingManner, ElegantMan
• Gemini Flash — Achernar, Achird, Algenib, Algieba, Alnilam, Aoede, Autonoe, Callirrhoe, Charon, Despina, Enceladus, Erinome, Fenrir, Gacrux, Iapetus, Kore, Laomedeia, Leda, Orus, Pulcherrima, Puck, Rasalgethi, Sadachbia, Sadaltager, Schedar, Sulafat, Umbriel, Vindemiatrix, Zephyr, Zubenelgenubi
• Gradium — English: Emma, Kent, Eva, Jack. German: Mia, Maximilian. Spanish: Valentina, Sergio. French: Elise, Leo. Portuguese: Alice, Davi

A voice not in the chosen model's list (and not a valid handle / custom Voice ID where allowed) → 400. An ElevenLabs Voice ID that the provider rejects also → 400 (not charged).

Voice cloning — POST /audio/voices

Clone a voice from an audio sample and get back a handle (vv_…) to pass as voice on /audio/speech. multipart/form-data only; max 25 MB.

tts-chatterbox-hd advertises its cloning contract on /models as model_spec.voice_cloning:

• Zero-shot: no voice template is derived; the reference audio is stored with a TTL and re-read on every synthesis call. Handles stop working 7 days after creation, regardless of use.
• mp4 covers M4A. Samples in other containers (or with a non-audio signature) → 400 before anything is uploaded.
• Each successful clone is charged a flat per-clone fee; synthesis is billed separately per character on /audio/speech.

tts-minimax-speech-02-hd also appears in the model enum in the OpenAPI spec, but cloning with it isn't open to regular keys, which get 403 "Voice cloning … is not available on your account". Its model spec on /models carries no voice_cloning object — use that as the signal.

A handle is bound to the model that created it. Pass it with a model that has no cloning support → 400; pairing it with a different cloning model fails.

Streaming

With streaming: true, the body is a chunked (Transfer-Encoding: chunked) audio stream — decode as it arrives. pcm (where supported: Kokoro, xAI, MiniMax, Gradium) is convenient for raw Web Audio playback. If the stream fails before headers are sent you get 500 {"error":"Stream error"}; after that the connection just ends.

OpenAI SDK

Style / emotion (Qwen 3, Gemini Flash)

tts-gemini-3-1-flash also takes prompt (style instructions) and temperature, but not top_p. For other families, delivery comes from the voice choice itself (e.g. Inworld Hades vs Pixie); prompt / temperature / top_p are silently ignored.

Errors

See venice-errors for body shapes and retry strategy.

Gotchas

• Always send model — the documented default never applies.
• input hard cap is 4096 chars. For long content, split on sentence boundaries and concatenate audio client-side.
• Don't assume mp3: Inworld, Chatterbox, Orpheus and Gradium default to wav, and any response_format outside a model's supported_formats returns 400. Omit response_format or check /models first.
• speed is only applied by Kokoro (unclamped), xAI, ElevenLabs Turbo and MiniMax (clamped). Keep 0.8–1.3 for natural narration.
• streaming is a Venice-specific field that isn't in the OpenAI SDK's types; pass it as an extra body field, or call the REST endpoint directly and consume the body.
• Voice names are case-sensitive (eve ≠ Eve, af_sky ≠ AF_SKY). Note leo (xAI/Orpheus) vs Leo (Gradium, French).
• Chatterbox cloned handles expire 7 days after creation. Re-clone rather than storing handles long-term.
• Gradium has no language parameter — pick the voice for the language you want.