Google dropped two new text-to-speech models yesterday: gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts. The standout feature isn’t the models themselves, it’s the library of over 2,000 voices they ship with, plus voice cloning from a 30-second sample.

Simon Willison built a playground for it in an afternoon using GPT-6 Astra. The Gemini API has open CORS, so you can build browser-based tools that hit it directly with your own API key. No server required.

This matters because voice cloning just got dramatically more accessible. Most TTS tools either give you a handful of generic voices or charge serious money for custom voice training. Google’s offering both a massive voice library and quick custom cloning in the same API.

The “flash-lite” variant is presumably faster and cheaper. The regular flash version trades some speed for quality. Google hasn’t published detailed pricing yet, but the pattern suggests this will undercut existing TTS services significantly.

What developers should know

The API is straightforward. You can start generating speech with a pre-built voice immediately, or submit a 30-second audio sample to create a custom voice profile. The legal requirement is explicit: you need rights to any voice you’re cloning.

Willison’s playground demonstrates the basic integration pattern. Since the API allows browser requests, you can prototype TTS features without backend infrastructure. That’s unusual for AI APIs and makes experimentation much faster.

The 2,000-voice library means you can probably find something close to what you need without custom training. But the 30-second cloning threshold is low enough that custom voices are practical for most projects.

The competition

ElevenLabs has been the go-to for realistic voice cloning, but requires more audio and costs more. Play.ht and Murf.ai offer similar features at various price points. OpenAI’s TTS API (part of GPT-5.6) is simpler but has a much smaller voice selection.

Google’s move is classic Google: take an established product category, add massive scale, price it aggressively. Whether the quality matches ElevenLabs or the API proves as reliable as OpenAI’s remains to be seen.

Who should care

If you’re building anything that generates speech (accessibility tools, content creation, voice assistants, audiobook production), this is worth testing. The combination of a huge voice library and easy custom cloning in one API is new.

If you’re already happy with your TTS setup and it’s working fine, there’s no urgent reason to switch. But if you’ve been avoiding voice features because existing tools were too expensive or complicated, the barrier just dropped.

The playground is live if you want to test it with your own Gemini API key.