Google’s Gemini 3.8 Flash TTS and Flash-Lite TTS models are generally available through the Gemini API. Flash targets more expressive speech generation, while Flash-Lite is positioned for higher-throughput, lower-latency workloads. Both support controllable speech generation, but they are audio-generation models—not complete conversational agents.
The two model IDs are gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts. Google marked them generally available in its API changelog on September 22, 2026, and published its overview the following day. The best choice depends on whether your product prioritizes expressive output or throughput, and both should be evaluated with representative text, accents, languages, and devices.
Flash vs. Flash-Lite: which model should you test?
| Model | Google’s positioning | Good first evaluation for |
|---|---|---|
| Gemini 3.8 Flash TTS | More expressive, creative speech generation | Branded narration, character dialogue, and content where delivery matters |
| Gemini 3.8 Flash-Lite TTS | Higher throughput and lower latency | High-volume announcements, short responses, and latency-sensitive workloads |
Those are provider descriptions, not a guarantee that one model will outperform the other in every application. Create a small test set from your actual scripts. Measure pronunciation, pacing, interruption tolerance, output latency, and the amount of post-processing your product needs.
What the API adds for voice design
The models accept text and speech-generation instructions, with controls for delivery such as voice and style. Google’s documentation lists a library of more than 150 voices through its Voices API, along with custom voice design and replication workflows. Google says its replication flow uses a short voice sample and includes consent verification.
Language support is not identical: Google’s model pages list 130 languages for Flash TTS and 101 for Flash-Lite TTS. Treat those counts as a starting point, not evidence of equal quality across languages. Test the names, dialects, code-switching, numbers, dates, and domain terms that matter to your users.
Important limits: TTS is not the agent layer
The model documentation lists important capabilities that these TTS models do not provide, including function calling and the Live API. In other words, they produce speech; they do not, by themselves, manage a real-time dialogue, authorize actions, search a knowledge base, or safely execute business workflows.
A production architecture may use a separate application or conversational model to decide what to say, then pass approved text to TTS. Keep policy enforcement and tool permissions in your application. Do not let generated voice output stand in for authentication, consent, or confirmation of a consequential action.
Pricing and rollout details to recheck
Google’s pricing page currently lists Flash-Lite TTS at $0.50 per million input tokens and $6 per million generated audio tokens through December 31, 2026; it lists $1 and $12 respectively beginning January 1, 2027. The page also expresses the lower introductory rate as approximately $0.0015 per 10 seconds of generated audio. Pricing can change, so confirm the live pricing page and your account’s billing before estimating production costs.
Consent, disclosure, and testing
Google describes consent checks for voice replication and watermarking/metadata measures for generated media. Those safeguards do not transfer responsibility away from the product owner. Obtain explicit rights to any source voice, explain when users are hearing synthetic speech, retain only audio samples you are authorized to process, and establish a deletion and incident process.
Before launch, evaluate voice quality with native speakers, measure end-to-end latency and cost, test retries and partial failures, and offer captions or text alternatives. A convincing synthetic voice should never make uncertain information sound verified or imply that a real person approved it.
FindMilan’s AI consulting service can help scope a speech-enabled product and its evaluation plan. The AI Web Awards project is a separate website-quality evaluation project, not a Gemini TTS deployment.
Official sources
Frequently asked questions
Are Gemini 3.8 Flash TTS and Flash-Lite TTS generally available?
Google's Gemini API changelog lists both text-to-speech models as generally available on September 22, 2026. Check the current model documentation and account access before deployment.
What is the difference between Gemini 3.8 Flash TTS and Flash-Lite TTS?
Google positions Flash TTS for more expressive, creative speech generation and Flash-Lite TTS for higher-throughput, lower-latency workloads. Benchmark both with your own prompts, languages, and latency targets.
How many languages do the Gemini 3.8 TTS models support?
Google's model documentation lists 130 languages for Gemini 3.8 Flash TTS and 101 for Flash-Lite TTS. Support does not guarantee identical pronunciation, quality, or latency in every language.
Can Gemini 3.8 TTS clone a voice?
Google documents custom voice design and voice replication from a short audio sample, with consent checks and usage restrictions. Only use voices when you have explicit rights and informed consent.
Can these TTS models call tools or act as live conversational agents?
No. Google's TTS model documentation describes speech-generation models and lists capabilities such as function calling and the Live API as unsupported. Use a separate orchestration or conversational model when those capabilities are required.