Google released two speech synthesis models on September 23: Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS. The former targets creative use cases that need fine-grained tone control and character work, while the latter is built for high-volume, low-cost calls. Both are available starting the same day in the Gemini API and Google AI Studio, with the enterprise-tier Gemini Enterprise to follow.
What it can do
Voices come from three sources: 30 built-in voices, more than 2,000 ready-made voices cleared for commercial use, and custom voices users create themselves. Cloning a custom voice takes just a 30-second audio sample, and Google says it supports more than 100 languages and dialects. Voice mixing is listed as "coming soon."
On the control side, developers can write natural-language performance directions line by line — telling the model where to lower its voice or add a laugh. The model natively supports two-speaker dialogue, and can insert scripted laughter, sighs, gasps, and filler responses like "mm-hm" and "right" on cue. Google also says audio quality holds up without noticeable degradation across several hours of generated content.
Both models are also reaching consumers directly: Flash TTS has landed in Gemini Notebook, and Flash-Lite TTS is now in Google Vids, both available to regular users. Ten launch partners — including Agora, LiveKit, Pipecat, Vercel, Figma, and HeyGen — are integrating the models.
Voice cloning comes with regional limits
Google built several safeguards around cloning: every generated clip carries a SynthID watermark, cloning requires consent verification beforehand, and outputs include C2PA content credentials. Cloning is not yet available in Illinois, Texas, the European Economic Area, the UK, Switzerland, or India. Illinois and Texas both have dedicated biometric information privacy laws.
On benchmarks, Google cites Hume AI's voice-design leaderboard, claiming Flash TTS ranks first overall and first for accent modeling; it also places near the top of Voice Arena's rankings for Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic, Mexican Spanish, and Hindi. These are third-party leaderboard results as relayed by Google — the announcement does not break out Chinese-language performance separately.
Running the numbers
Google's Gemini API pricing page counts audio output at 25 tokens per second, which works out to 90,000 audio tokens for an hour of speech. Rough cost per hour, based on price per million output tokens:
| Model | Through end of 2026 | From January 1, 2027 |
|---|---|---|
| 3.8 Flash TTS ($9 / $18) | ~$0.81/hour | ~$1.62/hour |
| 3.8 Flash-Lite TTS ($6 / $12) | ~$0.54/hour | ~$1.08/hour |
| 2.5 Flash preview TTS ($10) | ~$0.90/hour | — |
Text input costs $0.50 per million tokens through year-end and $1 after that — with a script running only tens of thousands of characters per hour, this cost is negligible. Batch processing gets a further 50% discount.
By that math, the new models undercut the old preview model for the rest of this year; starting next year, Flash pricing jumps about 80%, and Flash-Lite also ends up pricier than the old version. Google hasn't explained the "half-price now, double later" pricing structure. For audiobook, podcast, and customer-service workloads that burn through hours of speech, a cost of roughly a dollar an hour is still far cheaper than voice actors — audio quality and cloning fidelity will likely decide adoption more than a price swing of a few dimes.
Sources: Google's official blog, the Gemini API pricing page, CocoLoop, Hume AI; verified against voice counts, cloning sample length, regional restrictions, and per-million audio-token output pricing.