Gemini text to speech: four models, one decision

Four models, one script format

The current TTS line-up is two general-availability models and two previews. They accept the same kind of transcript, they draw from the same 30 preset voices, and they all stop at two speakers per request. If you learn one, you have learned the interface for all four. The differences show up in the audio, in the latency, and in which optional voice features you can reach.

The line-up at a glance

Status matters more than the version number. A general-availability model is the one to build a product on: behaviour is stable, quota is documented, and support exists when something breaks. A preview model is a moving target, and building a client deliverable on one is a scheduling risk rather than a technical one.

Which model for which job

The pattern that works for most teams is a two-stage loop. Iterate on Flash-Lite while the script is still moving: it returns fast enough that you can fix a sentence and hear the result in the same sitting. Once the script is frozen, render the final take on Flash. The audio is better, and because the script is stable you only pay for one pass.

What actually changes between generations

Three things, and they are worth understanding because they change how you write. First, the newer generation treats the transcript as a verbatim script: it says what is written, including stage directions, which is why acting notes moved into the style field. Second, delivery control is expressed as metadata rather than as prose in the text, so a single script can produce a calm narrator and an excited host without a rewrite. Third, vocal tags inside the line are handled more literally, which is a gain when you want a laugh in the right place and a liability when you leave angle brackets in by accident.

The older previews behave differently enough that a prompt written for the 3.8 generation can sound wrong on them. If you must keep a preview model in production, freeze its prompts alongside the model identifier and do not copy briefs across generations without an audition.

Latency, cost and the honest trade-off

Faster models are cheaper per token and return sooner, which matters when you are generating hundreds of clips for a catalogue, a course library or a localisation pass. Slower models are worth their cost in exactly one place: the take that ships. A useful habit is to budget two renders per finished minute — one on the fast model to lock the script, one on the fidelity model for the release.

Do not choose a model by its name. Generate the same eight-second line on each candidate and listen on the playback path your audience actually uses. Model preference arguments are almost always arguments about a specific sentence, and one sentence is not a benchmark.

Moving a project between models

• Re-audition the voice: the same preset name does not guarantee the same timbre across generations.
• Re-check the brief: style metadata is interpreted more literally by the newer models.
• Regenerate the tail of any long script and cross-fade it, rather than assuming the join still matches.
• Keep the model identifier in your project notes next to the voice name and the brief.

Frequently asked questions