Gemini 3.8 Flash TTS

Gemini 3.8 Flash TTS — Text-to-Speech API

What is Gemini 3.8 Flash TTS?

Gemini 3.8 Flash TTS is Google's flagship creative text-to-speech model, built to perform a script rather than just read it. Instead of picking a fixed preset, you supply a transcript, choose a studio voice, and direct the delivery in plain English — pacing, emotion, accent, and dialect all shift line by line. It is designed for developers and creators building audiobooks, podcasts, games, e-learning, and voice agents who need speech that sounds acted, not announced.

The model treats your text as a verbatim transcript and keeps all delivery control in a separate style field, so your prompt stays clean and predictable. It returns studio-quality WAV audio and supports automatic language detection across 130+ languages.

Key Features

• •Line-by-line style direction: steer tone, pacing, emotion, and accent with a natural-language style prompt.
• •Inline vocal cues: drop <laughs>, <sigh>, <gasp>, or <short pause> exactly where they belong in the transcript.
• •Native two-speaker dialogue: stage a two-person scene from a single call with distinct voices and natural turn-taking.
• •30 studio voices: expressive named voices such as Kore, Puck, Fenrir, Aoede, and Enceladus.
• •Multilingual: 130+ languages with authentic regional accents, auto-detected from the transcript.
• •Expressiveness control: a temperature dial from steady narration to animated performance.

Best Use Cases

Reach for Gemini 3.8 Flash TTS when delivery matters: character voices and emotional beats for games and immersive audiobooks, two-host podcasts and dramatic scenes, warm multilingual narration for explainers and localization, and natural-sounding voice agents. In testing, a single-voice sports call swung from a hushed build-up to a euphoric eruption, a two-speaker lab scene traded four turns in distinct voices, a Mexican-Spanish market read nailed the accent from the transcript alone, and the same engine dropped to an intimate whisper purely through the style field.

Prompt Tips and Output Quality

Keep stage directions out of text — it is spoken verbatim — and use the style field for emotion and the inline <angle-bracket> cues for sounds. Set voice_2 to a different voice for dialogue and prefix each line with the speaker name. Raise temperature to roughly 1.0–1.4 for expressive scenes; keep it near 0.3 for steady reads. Outputs are clean 24 kHz mono WAV.

How to Use Gemini 3.8 Flash TTS

Gemini 3.8 Flash TTS turns a written script into a performed voice track. The workflow is three clean parts: put what is said in text, pick who says it with voice_1 (and voice_2 for dialogue), and describe how it is said in style. Because text is read verbatim, never write directions like "say angrily" in it — put that in style, and use inline <angle-bracket> cues only for sounds.

Recommended Settings

• •text: your transcript, with inline cues such as <laughs>, <sigh>, <gasp>, <short pause> placed where they occur.
• •voice_1: one of 30 studio voices (Kore, Puck, Fenrir, Aoede, Enceladus, Callirrhoe...). Match the timbre to your character.
• •style: plain-English delivery direction applied to every line, e.g. whispered urgently or warm and inviting.
• •voice_2: set only for two-speaker scenes; choose a different voice and prefix each line with the speaker name.
• •temperature: 0.2–0.5 for steady narration; 0.8–1.4 for emotional, animated delivery.

Example voices to try

Each example below ran end to end at the settings shown, and the delivery direction carried over cleanly — emotion, accent, and whisper all came from the style field, not from the transcript.

• •Expressive emotion arc (default). Prompt: "It's fifteen seconds from the final whistle, and it's still Rovers one, City one. <short pause> Mateo Rinaldi picks it up on the halfway line... he's past one man, past two — can he finish it from here? <gasp> He shoots... and it's IN! Rinaldi wins it at the death, and the whole Olimpico is on its feet!" Voice Fenrir at temperature 1.1 swung from a hushed build-up to a euphoric goal call, with the pause and gasp landing inline. Tip: keep the emotional turn in style and the transcript clean.
• •Multilingual accent. Prompt: "¡Pásele, pásele! Estos son los mejores tacos al pastor de todo el mercado, se lo prometo. <short pause> Con su piña, su cilantro, su salsa verde... treinta pesos los tres. ¿Cómo los quiere, con todo?" Voice Callirrhoe delivered warm Mexican Spanish from the transcript alone — no language flag needed. Tip: write in the target language and keep inline cues in English.
• •Intimate whisper. Prompt: "The seal is intact. <short pause> Three thousand years, and no one has breathed this air. I'm easing the lamp through the gap now... <gasp> gold. Everywhere. The whole chamber is untouched." Voice Enceladus at temperature 0.9 produced a breathy, suspenseful whisper. Tip: lower temperature for controlled delivery and describe the voice quality in style.

Two-speaker dialogue

For a two-person scene, set voice_1 and voice_2 to different voices and start every line with that speaker's name. Example: "Aoede: Core temperature just hit forty-two degrees and it's still climbing. Are you seeing this? / Puck: Forty-two? <gasp> That's nearly double what the model predicted." Both voices stayed distinct with natural turn-taking in a single call.

Further scenes the model is built for: try "cheerful customer-service agent", "slow, calm bedtime story", or "breaking-news anchor, urgent".

Can I clone my own voice?

The Segmind API exposes 30 prebuilt studio voices rather than cloning; pick the closest timbre.

How many speakers can it do?

Up to two per call — set both voice_1 and voice_2.

What languages are supported?

Over 130, detected automatically from the transcript; keep inline cues in English.

How do I add laughs or pauses?

Write inline cues like <laughs> or <short pause> directly in the text.

What audio format does it return?

A standard WAV file (24 kHz, mono).

How is it different from Gemini 3.1 Flash TTS?

3.8 improves expressive acting, long-form stability, and two-speaker screenplay control.