Google's Gemini 3.8 Flash TTS Builds Custom Voices From Text Prompts

Google has added two text-to-speech models to its Gemini lineup, Gemini 3.8 Flash TTS and Flash-Lite TTS, both of which cover more than 100 languages. The headline capability is voice design: Flash TTS can construct a new synthetic voice from a written description rather than requiring a recording of an existing speaker. The pair arrived with little fanfare, but they push speech synthesis further from the pick-a-preset model that has dominated the category.

The two models split along a familiar cost-and-capability line. Flash TTS handles the heavier work, including generating voices that do not correspond to any real person, while Flash-Lite TTS targets lighter, faster jobs. Both accept stage directions attached to individual lines, so a script can specify delivery for one sentence without affecting the rest. The same script can also produce two-voice dialogue, letting a single document drive a back-and-forth exchange between speakers.

Voice cloning is included as well, and Google has set the input bar low: a 30-second sample is enough to build a voice profile. That figure matters because it puts cloning within reach of anyone with a phone and half a minute of quiet room, not just studios with professional recording setups. The combination of cloning and description-based generation means users can either copy a voice they already have or invent one that never existed.

The language count is the other number worth pausing on. Support for more than 100 languages puts the models in the same range as the largest commercial speech systems, and it is the kind of coverage that matters for dubbing, localization and customer support rather than for hobby projects. A single pipeline that handles triple-digit language counts lowers the cost of shipping the same product into many markets.

For the AI industry, the shift is from selecting a voice to specifying one. Description-based generation removes the licensing question that hangs over cloned voices, since the output is not tied to a real speaker, while the stage-direction feature moves expressive control from post-production into the script itself. Developers building audiobooks, games or voice agents gain finer control without stitching together multiple tools.

What remains open is how Google will police the cloning side. A 30-second sample is a low bar, and the company has not detailed consent checks, watermarking or restrictions on whose voice can be reproduced. How those guardrails are enforced, and whether Flash-Lite TTS reaches the same language coverage as its larger sibling, will decide how quickly developers move their speech stacks over.