Dialogue tools

Generate speech with the generate_dialogue tool. Pick one of four providers with the provider parameter:

• ElevenLabs v3 (provider: "elevenlabs", the default): 26 human-named voices, the most natural prosody. One voice per call.
• Google Gemini 3.1 Flash TTS (provider: "google" or "gemini"): 30 voices, 70+ languages, a natural-language style prompt, and inline expressive tags ([sigh], [laughing], [whispering], [shouting], [extremely fast], [like dracula]) that are actually performed. One voice per call.
• Gemini 3.8 Flash TTS (provider: "gemini-3.8"): a two-speaker conversation in one call, with a delivery style per line and angle-bracket tags such as <laugh>. 70 voices.
• Seed 1.0 Audio (provider: "seed"): prompt-driven generative audio. You describe the voices, sound effects and music in prose instead of picking a voice, and can clone up to 3 voices from audio references.

Rates differ by provider and are never fixed in these docs: read them from get_models (filter model_type: "speech") or call estimate before generating. ElevenLabs, Gemini 3.1 and Gemini 3.8 bill per 1,000 characters; Seed bills per second of output.

To dub an existing video into other languages, use generate_dubbing below.

#ElevenLabs v3 (default)

elevenlabs/v3 accepts these 26 voices (case-sensitive). Any value outside this list is rejected with 400 Unsupported voice and you are not charged. Audio tags like [laughs] are stripped server-side before synthesis (not spoken).

#Female voices

#Male voices

#Google Gemini 3.1 Flash TTS

Pass provider: "google". Gemini adds a natural-language style prompt (set the tone, pace, accent, emotion, or a character) and performs inline [tags] instead of stripping them. It accepts 30 voices (case-sensitive) and BCP-47 language codes.

Inline tags are performed, not stripped: [sigh], [laughing], [uhm], [whispering], [shouting], [sarcasm], [robotic], [extremely fast], [short pause] / [medium pause] / [long pause], and free-form descriptive tags like [like dracula] or [excitedly].

#Female voices

#Male voices

Languages: BCP-47 codes (e.g. en-US, en-GB, es-ES, es-MX, pt-BR, fr-FR, de-DE, it-IT, ja-JP, ko-KR, hi-IN, ar-001, ru-RU, ro-RO, tr-TR, vi-VN, th-TH, and 70+ more), or auto to auto-detect. Default en-US. Call get_limits_for_model('text-to-dialogue-gemini') for the full list.

<a id="gemini-3-8"></a>

#Gemini 3.8 Flash TTS

Pass provider: "gemini-3.8". This is the provider for a conversation between two voices, or for one voice whose delivery changes line by line. Output is WAV (24 kHz mono).

• Simple: text + voice (default Fola), with style_prompt as that line's delivery.
• Dialogue: speakers is a list of 1 or 2 items, [{"voice": "Fola"}, {"voice": "Zephyr"}], numbered Speaker 1 and Speaker 2 in order. dialogue_turns is the ordered script, [{"speaker": "Speaker 1", "text": "...", "style": "slow and dry"}, ...], up to 50 turns. Every turn's speaker must be declared in speakers. text is ignored when dialogue_turns is set.
• style per turn (optional): free-form tone, emotion or pace, for example "tired night-shift radio host, slow and dry". Up to 500 characters per turn and 2,000 across all turns.
• filler_words: true adds natural hesitations such as um and uh. It only applies with two speakers.
• Tags use angle brackets, inline in a turn's text: <laugh>, <sigh>, <whispers>, <short pause>, <long pause>, <gasp>, <chuckle>, <breath>, <cough>, <yawn> and the rest of the 40 supported tags. [square] tags are not performed by this model.
• No language_code: the model reads the language from the text.
• Limit: up to 5,000 characters across all turns, tags included. Unknown voices or speakers are rejected before billing.
• Billing: per 1,000 characters of turn text, tags included, with a minimum of 1 credit.

Voices (70), grouped by use case:

Call get_limits_for_model('gemini-3-8-flash-tts') for the current list.

<a id="seed"></a>

#Seed 1.0 Audio

Pass provider: "seed". Seed is prompt-driven: text is a free-form "director's script" that can produce voices, sound effects and background music together, not a fixed voice reading words verbatim. There is no voice parameter, [tags] are not used, and the ElevenLabs tuning parameters do not apply. Describe each voice (gender, age, accent, timbre, emotion) in the prose. Fetch the seed-audio-prompting skill before writing the prompt; a bare line gives bare output.

• text (required): the prompt, up to 2,048 characters.
• References (optional; image and audio cannot be mixed):
  • one image, as image_url (HTTPS) or base64 image_data, to guide voice and mood; or
  • up to 3 audio references for voice cloning, as audio_references: a list where each item is one of {"audio_url": "..."}, {"audio_data": "<base64>"} or {"speaker": "..."}, each clip up to 30 seconds. Tag each cloned voice in text by upload order as <<TGT_SPK1>>, <<TGT_SPK2>>, <<TGT_SPK3>> (for example Marcus (..., the actor is <<TGT_SPK1>>) says: "..."). @Voice1, @Audio1 and @Speaker1 are also accepted and normalized server-side.
• audio_config (optional): format (mp3, wav, ogg_opus, pcm), sample_rate, speech_rate (-50 to 100), loudness_rate (-50 to 100), pitch_rate (-12 to 12).
• Limits: each reference up to 10 MB, at most 3 audio references or 1 image, output capped at 120 seconds.
• Billing: per second of output (capped at 120 seconds), minimum 1 credit.

Example (Seed, two voices described in prose):

{
  "provider": "seed",
  "text": "A quiet coffee shop at dawn, soft rain on the windows. A warm, low-voiced barista in her fifties says: \"First cup's on the house.\" A sleepy young man laughs and answers: \"You say that every morning.\""
}

#generate_dialogue

Generate a speech clip.

Parameters:

ElevenLabs strips [HEY] / [laughs] / [whispers] before synthesis. Google performs inline tags — see the tag list above.

Example (Google Gemini):

{
  "provider": "google",
  "text": "[whispering] I have a secret. [laughing] Just kidding!",
  "voice": "Callirrhoe",
  "style_prompt": "Speak playfully, like sharing a fun secret with a friend.",
  "language_code": "en-US"
}

Example (Gemini 3.8, two speakers):

{
  "provider": "gemini-3.8",
  "speakers": [{ "voice": "Fola" }, { "voice": "Zephyr" }],
  "dialogue_turns": [
    { "speaker": "Speaker 1", "text": "Did you hear that? <short pause> Something moved.", "style": "nervous, whispering" },
    { "speaker": "Speaker 2", "text": "<laugh> It's the cat. It's always the cat.", "style": "amused, relaxed" }
  ],
  "filler_words": true
}

Example (ElevenLabs):

{
  "text": "Welcome back, ready to generate?",
  "voice": "Rachel",
  "stability": 0.5,
  "language_code": "en"
}

Response: Returns a generation_id. Poll with get_status.

#Multi-voice scenes

For a conversation between two speakers, use Gemini 3.8 (provider: "gemini-3.8") with speakers and dialogue_turns: the whole exchange comes back as one file. Seed can also voice several characters in one prompt. ElevenLabs and Gemini 3.1 are single-voice: call the tool once per line (ElevenLabs: pass previous_text / next_text for prosody continuity), then concatenate the resulting audio files yourself.

#get_status

Use the generation_id returned by generate_dialogue to check progress. When the status is completed, the audio file URL is in the outputs array (media_type: "audio").

<a id="generate_dubbing"></a>

#generate_dubbing

Translates and re-voices the speech already in a video, and returns one dubbed MP4 per language. Two engines: sonilo-dubbing (default) and elevenlabs-dubbing (ElevenLabs, 90+ languages including Romanian, voice cloning, a lower per-second rate). It writes no music, effects or subtitles, and it cannot add speech to a silent clip. This is the most expensive model on the platform and generations cannot be refunded, so quote with estimate and get the user's explicit go-ahead first.

Parameters:

Billing: clip seconds multiplied by the number of languages, at the engine's own per-second rate. Sonilo applies a 10-second minimum to that total rather than to each language; ElevenLabs has no minimum. Three languages cost three times one. Quote it with estimate(model=<engine id>, reference_video_seconds=<clip seconds>, variant_count=<number of languages>).

Response: {generation_id, status: "pending", languages}. Poll get_status until completed or failed: the poll is also what settles the run, so keep polling to the end. Each output's metadata.language names its language.

#Credits and limits

• Cost: read the live rate from get_models or call estimate; rates differ by provider. ElevenLabs, Gemini 3.1 and Gemini 3.8 bill per 1,000 characters (rounding: decimal ≤ 0.3 floors, > 0.3 ceils; Gemini 3.8 counts tags as characters). Seed bills per second of output, capped at 120 seconds.
• Minimum: 1 credit for any non-empty request.
• Minimum length: 5 characters after audio tags are stripped.
• Maximum length: ElevenLabs 5,000 characters; Google 4,000 bytes for text (and ≤ 8,000 bytes for text + style_prompt combined); Gemini 3.8 5,000 characters across all turns, tags included; Seed 2,048 characters of prompt.

See Limitations for full details.