Gemini 3.8 text-to-speech

TL;DR Highlight

Gemini 3.8 Flash TTS designs voices from prompts and directs delivery line by line.

Who Should Read

A backend developer choosing a TTS API for voice agents, audiobooks, or dubbing.

Core Mechanics

Two TTS models, two jobs
- Gemini 3.8 Flash TTS is for deep creative direction and character design.
- Gemini 3.8 Flash-Lite TTS is for high-volume dubbing, content, and voice agents.
Design voices from a prompt
- Natural language sets role, accent, and voice traits across 100+ languages and dialects.
- The library has 2,000+ production-ready voices.
Clone a voice from 30 seconds
- A 30-second sample of your voice, or one you have rights to, is enough.
- Consent verification, SynthID watermarking, and C2PA credentials are built in.
Direct each line
- You control acting cues, pacing, dialect shifts, and backchanneling (listener interjections).
- Cues like <laughs>, <sigh>, <gasp> and |mhm|, |yeah| go right in the script.
Long audio and two speakers
- It keeps quality across hours of audio with minimal speaker drift.

Community Reactions

Opinion: prompts get ignored
- Some said core parts of the prompt are ignored in most demos.
- Another noted the Monologue demo skips cues like <chuckles> and <laughing>.
Swapped in for Eleven v3
- One user replaced Eleven v3 TTS in an internal listening feature and found expressiveness on par.
- They said integration was trivially easy.
Cost and rollout reactions
- One user found most experiments in conversation mode cost under a cent.
- Another hit an error in Voice Design and complained about missing pricing.
Opinion: platforms are inconsistent
- Availability and capabilities differ across Google's consumer, prosumer, and cloud platforms.
- This can block use when an org disables some of them.
A local alternative exists
- One user built a local audiobook tool with Gemma 4 and Qwen3 TTS Voice Design.

How to Apply

Run your own script in the AI Studio playground
- Today, paste 5-10 real lines into the dual-speaker editor.
- Add per-line direction and cues like <laughs>, then check what is honored.
A/B listen against your current TTS
- Generate the same script with your current TTS, Flash TTS, and Flash-Lite TTS.
- Compare expressiveness, cue handling, and speaker consistency.
Start with Flash-Lite for volume
- Apply Flash-Lite TTS to your bulk dubbing or voice agent paths first.
- Upgrade only weak lines to Flash TTS.
Check region and consent before cloning
- Check whether you serve Illinois, Texas, EEA, UK, Switzerland, or India.
- If so, AI Studio voice replication is unavailable, so pick an alternative.

Try It

copy & run

# Paste into Claude or ChatGPT (build a line-by-line direction script for Gemini 3.8 TTS, simplified version of the article's method)

You are a TTS director. Turn the text below into a two-speaker script.
- First write one natural-language line describing each voice for speakers A and B (role, accent, manner).
- Put an acting direction in parentheses before each line (tone, pace, emotion).
- Only where needed, add non-verbal cues like <laughs>, <sigh>, <gasp> and backchannels like |mhm|, |yeah|.

{{source dialogue or description}}

Next step: paste the result into the dual-speaker editor in the Google AI Studio audio playground and check which cues are honored.

Terminology

TTSA model that reads text aloud as speech. It automates handing a script to a voice actor and getting a recording back.

backchannelingListener interjections like 'mm-hm' or 'yeah' during someone else's speech. They make dialogue sound natural.

SynthIDAn inaudible watermark embedded in AI-generated audio. It lets you detect later that the audio is AI-made.

C2PAA standard for credentials proving a file's origin and edit history. Think of it as a résumé attached to the file.

speaker driftWhen a speaker's voice gradually changes over long audio.