Gemini TTS Explained: 3.1 Flash Preview Voices, Pricing, and Limits (2026)
Last updated August 7, 2026
Gemini TTS is Google's preview-status controllable speech line: gemini-3.1-flash-tts-preview plus the 2.5 Flash/Pro TTS predecessors. It offers 30 prebuilt voices with personality labels, natural-language style direction with inline tags like [whispers], 90+ languages with auto-detection, a 32k-token context — and a two-speaker cap despite the multi-speaker branding. No voice cloning at all, which compliance-minded teams can treat as a feature. Runs in invideo's Audio tab.
Updated August 2026
Gemini TTS is Google's controllable text-to-speech line in the Gemini API — current model gemini-3.1-flash-tts-preview, with 2.5 Flash and Pro TTS as predecessors — and the first thing to know is that every model in the line is still in preview as of August 2026: no SLA, tighter rate limits, specs that can change. What it offers is distinctive anyway: 30 prebuilt voices you direct with natural-language instructions ("speak like a tired night-shift radio host"), inline audio tags like [whispers], support across 90+ languages with automatic detection — and no voice cloning at all, which for some teams is precisely the point.
Which Gemini TTS models exist?
Per the official speech generation docs, August 2026:
| Model | Status | Notes |
|---|---|---|
| gemini-3.1-flash-tts-preview | Preview | Current model; free tier available |
| gemini-2.5-flash-preview-tts | Preview | Cheaper predecessor |
| gemini-2.5-pro-preview-tts | Preview | Prior quality tier; no free tier |
Preview status is not a footnote. Google's docs note preview models carry "more restrictive rate limits," and preview models can change or be withdrawn — don't build a launch-critical pipeline on an SLA that doesn't exist.
Gemini TTS specs at a glance
| Spec | Detail (official docs, Aug 2026) |
|---|---|
| Prebuilt voices | 30, with personality labels — Kore "Firm," Puck "Upbeat," Enceladus "Breathy," Sulafat "Warm" |
| Speakers per generation | 2 maximum, via speaker_voice_config |
| Languages | 90+ with automatic language detection |
| Direction | Natural-language style prompts + inline audio tags ([whispers], [excitedly], [very slow]) |
| Context window | 32k tokens per session |
| Modality | Text in, audio out only |
| Voice cloning | None — preset voices only |
How does multi-speaker mode work — and what's the catch?
The catch is the number: despite the "multi-speaker" branding, Gemini TTS generates at most two speakers per request. You assign each speaker a name and a prebuilt voice through speaker_voice_config, and the model matches names in your script to voices. That covers interviews, dialogue scenes, and host-and-guest formats — but a three-character scene needs multiple generations stitched in the edit. The cap is documented, not a bug, and it's the most common surprise in the line.
The upside is coherence: both voices generate in one pass, so pacing and turn-taking sound like an actual exchange rather than two files butted together.
How do you direct a Gemini TTS voice?
This is the line's real differentiator: direction happens in plain language rather than parameter soup. Three documented layers stack:
- Style prompts — prepend an instruction: "Read this in a calm, reassuring tone, slowing down for the final sentence."
- Inline audio tags —
[whispers],[excitedly],[very slow]placed exactly where the delivery should change. - Directorial prompting — the docs describe an Audio Profile / Scene / Director's Notes structure for building a full performance brief around the script.
Because the same 32k-token context carries both instructions and script, you can direct long passages with scene-level notes the way you'd brief a voice actor — no SSML required.
How much does Gemini TTS cost?
Official Gemini API pricing, August 2026:
| Model | Input (text) | Output (audio) | Free tier |
|---|---|---|---|
| Gemini 3.1 Flash TTS Preview | $1.00/M tokens | $20.00/M tokens | Yes |
| — batch mode | $0.50/M tokens | $10.00/M tokens | — |
| Gemini 2.5 Flash TTS | $0.50/M tokens | $10.00/M tokens | No |
| Gemini 2.5 Pro TTS | $1.00/M tokens | $20.00/M tokens | No |
One comparability warning: this is token pricing, not character pricing. Most TTS vendors quote dollars per million characters of script; Google bills per million tokens, and most of the cost is output audio tokens, which scale with audio length rather than script length. A "$X/M chars vs $20/M tokens" comparison is apples to oranges — benchmark your own script-to-cost ratio on the free tier before committing.
Why is there no voice cloning?
Gemini TTS ships preset voices only — you cannot upload a sample and clone a voice, full stop. For creators wanting their own voice, that's the honest reason to pick a different engine. But for compliance-minded teams it reads as a feature: no cloning surface means no consent verification to administer, no impersonation exposure, and no risk of a contractor uploading a voice they had no right to use. A workflow built on Gemini TTS is structurally incapable of producing an unauthorized voice clone — a property no cloning-first TTS family can offer.
What can you make with Gemini TTS?
- Multilingual narration — 90+ languages with auto-detection make it a strong engine behind an AI voice generator workflow when one script needs many locales.
- Dialogue scenes — two-speaker generation in a single coherent pass suits interview formats and explainer dialogues, within the two-voice cap.
- Course narration — directable pacing ("slow down for the definition") and consistent preset voices fit educational video production, where a stable voice across dozens of lessons matters more than cloning.
Gemini TTS FAQ
Is Gemini TTS generally available?
No. All three models — gemini-3.1-flash-tts-preview and both 2.5 TTS models — are in preview as of August 2026, with tighter rate limits and no GA guarantees, per Google's official docs.
How many speakers can Gemini TTS generate at once?
Two, maximum, via speaker_voice_config — despite the "multi-speaker" branding. Scenes with three or more characters require multiple generations.
Can Gemini TTS clone my voice?
No. The line offers 30 prebuilt voices only, with no cloning capability of any kind. Voice identity comes from choosing and directing a preset.
How much does Gemini TTS cost?
Gemini 3.1 Flash TTS Preview: $1.00 per million input text tokens and $20.00 per million output audio tokens, with a free tier and half-price batch mode (official pricing, August 2026). Note this is token-based pricing, not directly comparable to per-character TTS rates.
What languages does Gemini TTS support?
The docs list support across 90+ languages — including English, Spanish, Mandarin, Japanese, Arabic, and Hindi — with automatic language detection.
Where can you use Gemini TTS?
You don't need a Google AI Studio key to hear it: the invideo agent surfaces the line in its Audio tab as "Gemini 3.1 Flash TTS — multi-speaker text-to-speech with 30 voices," so directing a preset voice over your footage happens where the rest of the edit does. And because every model on the platform shares one workspace, a Gemini-narrated lesson can pull its visuals from the video models next door in the same project.
Version history: gemini-2.5-flash-preview-tts and gemini-2.5-pro-preview-tts — 2025 preview line → gemini-3.1-flash-tts-preview — current model with free tier, 30 voices, 90+ languages. All models remain preview-status. Facts current as of August 2026, per Google's official Gemini API documentation and pricing pages.