Introduction

Google just turned Gemini TTS from a list of presets into a full voice studio. On September 23, 2026, the company launched Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS—two text-to-speech models aimed at creators who want to design a voice, then direct every line.

This is a different product surface from Gemini 3.8 Live (the speech-to-speech agent path we covered earlier). TTS here is for exact scripts: audiobooks, podcasts, games, dubbing, and agents that need controllable recitation—not free-form multimodal chat audio.

Pro Tip

Treat the two models as a creative/cost pair. Use gemini-3.8-flash-tts when acting, dialect shifts, vocal bursts, and long-form stability matter. Reach for gemini-3.8-flash-lite-tts when you need high-volume dubbing, read-aloud, or agent cascades at lower cost. Same API shape; one parameter swap.

What Shipped Today

Google’s official blog and the Gemini API docs both date the drop Sep 23. Trade coverage (The Next Web, ~15:58 UTC / ~23:58 MYT) matches the same-day rollout.

Where you can try it now:

  • Developers: Gemini API + Google AI Studio audio playground
  • Flash TTS extras: rolling into Gemini Notebook
  • Flash-Lite TTS: Google Vids for everyone
  • Enterprises: Gemini Enterprise API coming soon

Model IDs in the developer docs: gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts. Docs note last update 2026-09-23 UTC.

Design a Voice, Don’t Just Pick One

Flash TTS’s headline move is generative voice design: describe role, accent, and character traits in natural language and get a bespoke vocal persona—Google demos range from a high-energy Melbourne DJ to a Japanese dragon. Coverage spans more than 100 languages and dialects on the marketing side; the API docs list 130 languages for Flash TTS and 101 for Flash-Lite.

You also get:

  • 2,000+ production-ready voices in an extended library (regional examples called out: Mexican Spanish, Quebec French, Scots English), plus 30 curated studio voices (Kore, Puck, Zephyr, and friends)
  • Voice replication from about 30 seconds of reference audio, with a verbal consent check that the consent speaker matches the reference
  • Save and reuse custom voices across projects (API: prompted/replicated voice_… IDs; stateful voices up to 200 per project, ~1-year TTL)
  • Voice remixing (timbre, pitch, pace, accent via prompts)—coming soon

Safety is not a footnote: every Gemini Audio clip carries SynthID watermarking, and Google mentions C2PA credentials for transparency. Voice replication through AI Studio is not available in Illinois, Texas, the EEA, UK, Switzerland, and India.

Direct the Performance Line by Line

Both models let you steer delivery without stuffing stage directions into the spoken transcript. Google’s 3.8 prompting model splits control:

  • Turn-level style via speech_metadata (emotion, pace, whisper, sarcasm)
  • Point-in-time vocal bursts as inline angle-bracket tags (laughs, sighs, pauses)
  • Native two-speaker scenes from one script, with conversational turn-taking
  • Backchanneling with pipe reactions (|mhm|, |yeah|) for overlapping texture
  • Long-form generation aimed at hours of podcast/audiobook audio with less speaker drift

On Hume AI’s Voice Design Benchmark, Google claims Flash TTS at #1 overall (71.4) and leading accent modeling (60.8), with Flash and Flash-Lite taking #1 / #2 on Hume’s Overall Quality Index. Blind Voice Arena prefs are highlighted for Japanese, Brazilian Portuguese, Vietnamese, MSA Arabic, Mexican Spanish, and Hindi.

How Builders Call It

The Interactions API path is text-in, audio-out. Unary requests default to WAV (audio/wav, 24 kHz mono). Streaming returns headerless L16 PCM chunks you can play as they arrive. Multi-speaker mode configures speakers in speech_config and tags each turn with a speaker in metadata.

Partners already named for integrations include Agora, LiveKit, Pipecat, Vercel on the infra side, plus product teams like Figma, HeyGen, Linguana, Wondercraft, 99.co, and Ollang for dubbing and agents.

If you are migrating from gemini-3.1-flash-tts-preview, move sustained directions out of the transcript into speech_metadata, keep only moment tags inline, and expect WAV by default on unary calls instead of raw PCM.

How This Fits the Gemini Audio Family

Google positions 3.8 Flash TTS beside earlier Gemini Audio pieces: 3.5 Live Translate, 3.5 Transcribe, 3.8 Live, and 3.8 Live Extended Thinking. Keep the lanes clear:

SurfaceJob

3.8 Flash / Flash-Lite TTS

Exact script → controllable speech

3.8 Live

Interactive speech-to-speech agents

Lyria (separate music stack)

Music generation, not dialogue TTS

ArtRealmAI already covered Live agents and Lyria music; this article is the scripted TTS studio drop.

Conclusion

Gemini 3.8 Flash TTS is the creative director; Flash-Lite is the volume knob. Same day as the launch, you can open AI Studio, prompt a new character voice (or consent-check a clone), and stage a two-speaker scene with line-level acting cues. For voice creators, podcasters, game writers, and agent builders, that is the difference between “pick a preset” and “cast and direct.”

Bookmark the Google blog post and the Gemini API speech-generation docs, then try a voice design pass before you lock a production narrator.

—Anabel ♡