Google’s Gemini 3.8 AI voices: what you can actually use

Google has announced two new speech-generation models, Gemini 3.8 Flash TTS and Flash-Lite TTS. The 23 September launch is aimed at turning a written script into directed audio, including custom voices and dialogue.

What does text-to-speech actually mean here?

Google’s developer guide distinguishes TTS from its Live API. TTS takes written text and produces audio, with controls for delivery such as accent, pace and tone. The Live API is intended for interactive conversation. If your goal is a narrated explainer, start by looking at TTS; if it is a voice assistant that responds to interruptions, these are different requirements.

A sensible use would be a short product walkthrough whose wording is already settled. Keep the script as your source of truth. The ability to make a sentence sound convincing should not decide whether the sentence belongs in the recording.

Two speakers does not mean every voice combination works

The API documentation limits single-request multi-speaker generation to two speakers using prebuilt voices. For dialogue using custom-designed or replicated voices, it describes generating each speaker’s turn separately and joining the audio. That extra work matters if you are imagining a one-click, fully customised two-host podcast.

There is also a size limit to plan around. Google’s model card lists up to 8K tokens of text input and 64K tokens of audio output for the TTS models. Output tokens are not words, and the two figures are not interchangeable. For a long script, check how your chosen tool handles sections before building a workflow around a single upload.

Voice copying has consent checks

Google says voice replication requires a verbal consent recording matching the voice owner. It also says generated audio carries its SynthID watermark. Voice remixing is described as coming soon. These are the company’s stated safeguards and roadmap, not a guarantee that misleading audio is impossible.

A useful first test for an Australian script

Before producing a long recording, try a small sample that is awkward in the ways your real material will be awkward. Include an Australian place name, an acronym, a price and a corrected date. For example: a Wollongong appointment, an NBN connection, twenty-nine dollars and ninety-five cents, and a booking moved from Friday to Thursday.

Listen while following the script, then listen once more without looking. Did the price remain unambiguous? Can you understand the acronym? Is Thursday clearly the new date? Finally, replace one sentence and check whether the new take fits beside the old one. These are suggested evaluation steps; we have not performed them on the new models.

Do that small test before paying for a larger production workflow. Also check the access and billing terms of the particular app or API you choose. The rollout announcement alone is not evidence that your subscription includes every voice feature.

For an explainer, clarity is a better starting point than a dramatic voice. If the recording makes the instructions harder to follow, a more expressive performance has not improved the result.