Guides

Gemini Voice Design and Remix Prompts: Templates for Consistent AI Voices

Prompt templates for Gemini 3.8 TTS voice design and remixing: describe a voice, change one trait at a time, and keep per-line style short to avoid drift.

Google’s Gemini 3.8 text-to-speech models can create a voice from a written description (voice design) and, since November 4, 2026, change an existing voice with plain-language instructions (voice remixing). Both run in Google AI Studio and the Gemini API with Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS (news summary). The facts below come from Google’s speech docs; the templates are our own. Results vary, so listen to the preview before you commit.

How the pieces fit

  • Voice design: a description in, a stored voice with a voice_... ID and a WAV preview out.
  • Voice remix: a base voice (prebuilt or designed) plus change instructions in, a new independent stored voice out. The original stays as it was.
  • Style per request: a short note on delivery (pace, emotion, volume) for one line or turn. The text itself stays the exact script.

Google’s main advice: put the persona in the voice, not in a long style block on every request. Long “director’s notes” are, in Google’s words, a common cause of voice drift.

Template: design a voice

A [age range] narrator with a [low / medium / high] pitch and a [warm / crisp / husky] tone. Speaks at a [slow / steady / brisk] pace with [clear / relaxed] diction and a [neutral / regional] accent. Sounds [friendly and reassuring / confident and direct / curious and playful]. Suited to [audiobook chapters / product explainers / a kids' science show].

Keep it to the traits that will not change from line to line. Leave out names and anything that points to a real person.

Examples you can adapt

A mature narrator with a low, warm tone and an unhurried pace. Clear diction, gentle pauses at the end of sentences, calm and kind. Suited to bedtime stories.
A young adult presenter with a bright, crisp tone and a quick, steady pace. Smiles while speaking, sounds sure of the facts without hype. Suited to 30-second app tutorials.
A middle-aged voice with a husky, relaxed tone and a slow, conversational pace, like someone talking across a kitchen table. Suited to a cooking podcast intro.

Template: remix one trait at a time

Remixing works best as a small, specific change. Name what to change and what to keep.

Keep the same voice, tone and accent. Make the pace about 20% slower and add slightly longer pauses between sentences.
Keep the pitch and pace. Make the tone warmer and softer, less formal, as if speaking to one friend rather than an audience.
Keep everything the same, but make the energy higher and more upbeat for short promotional reads.

Change one thing per remix and name each result clearly (for example “narrator-slow”, “narrator-warm”), so you can go back to the version that worked. Each remix is a separate stored voice, and Google caps a project at 200 stored voices, so delete experiments you do not need.

Short style notes that do not cause drift

Use the per-request style for momentary delivery only:

quietly, as if sharing a secret
slightly faster, excited
slow and serious

Google’s docs say not to put age, gender, names or permanent accent changes in the style, and not to add instructions like “keep the voice steady”, since extra text can make drift worse. For short sounds, Google supports inline tags in angle brackets in the script, such as <short pause>, <sigh> or <laugh>, and says to use human sounds rather than sound effects.

Dialogue with two or more designed voices

A single multi-speaker request supports up to two speakers using prebuilt voices only. To use designed or remixed voices in a dialogue, Google says to generate each speaker’s turn separately and join the audio. Keep a simple script sheet: line number, voice ID, text, style note.

For voice agents

Google suggests Gemini 3.8 Flash-Lite TTS for high-volume work such as voice agents, one TTS call per turn, and either no style or one short constant string, without resending a long persona each turn. For the conversation side, see how to write voice agent prompts.

Rights and safety

Only design or remix voices you have the right to use, never try to imitate a real person without their permission, and tell listeners a voice is AI-generated where the law or your policy requires it.

More voice ideas: AI voice prompts, Eleven v4 audio tags, the Gemini tool page and the model tracker.

stay in the loop

New guides, prompts and AI model news by email