gemini-3.1-flash-tts

Prerequisites

Install and load these skills before generating (skip if already in context via @pruna):

Or install the full suite once: npx skills add PrunaAI/pruna-skills@pruna -y

Follow each skill's Before generating / craft sections — do not restate guide content here.

Agent habit

In the first reply, name `gemini-3.1-flash-tts` in backticks, confirm REPLICATE_API_TOKEN (or stop with signup links from pruna-api), then ask for required inputs. Open intake → generation-diversity clarification intake (locale, voice, script) before the first POST. Redirect when When NOT to use fits better.

When NOT to use

Use a different skill instead:

Environment

export REPLICATE_API_TOKEN=r8_...

Requires ffmpeg / ffprobe when trimming, concatenating scene VO, or mixing with a bed.

HTTP (curl)

curl -s -X POST \
  -H "Authorization: Bearer ${REPLICATE_API_TOKEN}" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "text": "[warmly] The plush went flying. [short pause] And then it was gone.",
      "voice": "Sulafat",
      "prompt": "Warm storybook narrator, gentle pace, empathetic, no announcer voice.",
      "language_code": "en-US"
    }
  }' \
  "https://api.replicate.com/v1/models/google/gemini-3.1-flash-tts/predictions"

Poll urls.get until status is succeeded; download output (audio URL). Shared client: follow pruna-api (Replicate HTTP in the tool skill).

Before generating

1. Complete Prerequisites guide reading order.
2. Confirm text, voice, prompt, and language_code with the user. text, prompt, and inline [tags] must align — same emotional direction (see audio-prompting tts-style-prompting). When listing fields, name REPLICATE_API_TOKEN (Replicate — not PRUNA_API_KEY).
3. Model notes: combined text + prompt ≤ ~8,000 bytes; output capped ~655s. When TTS feeds p-video as input.audio, keep each line ≤ ~19s (ffprobe) — P-API clips audio at 20s. Common voices: Kore, Aoede, Sulafat, Achird, Charon, Puck, Vindemiatrix — full list on the Replicate readme.

Required input

• text (string) — spoken copy; supports inline [tags]. Max ~4,000 bytes.

Common optional fields

• voice (default Kore)
• prompt — style / director notes (max ~4,000 bytes)
• language_code — BCP-47 (default en-US)

Inline tags (examples): [sigh] [laughing] [whispering] [short pause] [medium pause] [long pause] [excitedly].

Typical next steps

Common follow-ons after this skill: