Google has released two speech models that let users invent a voice from a written description or copy a consenting adult’s voice from a short recording. Gemini 3.8 Flash TTS and Flash-Lite TTS combine those features with introductory audio-output charges of 81 cents and 54 cents per generated hour, respectively, putting custom narration and synthetic dialogue within easier reach of small publishers and creators.
Both are generally available through the Gemini API, which developers use to connect models to their software, and Google AI Studio, its browser-based workspace. Google’s launch announcement says rollout began September 23, although its release notes list general availability under September 22.
The hourly figures below are calculated from Google’s published standard rates. They cover generated audio only, before text input or other production costs. The introductory prices run through December 31, 2026.
At those introductory rates, ten generated hours would carry audio-output charges of $8.10 on Flash or $5.40 on Flash-Lite. That makes experimenting with audiobook narration, dubbing or podcast dialogue inexpensive, though a finished production still needs editing and review. Both models also have free-tier access; Google’s pricing page says free-tier content may be used to improve its products, while paid-tier content is not.
There are two distinct ways to create a custom voice. Voice design turns a text description into a reusable vocal persona. Voice replication uses a recording to reproduce a particular speaker’s vocal characteristics. Both models support one- or two-speaker output. Google’s speech guide lists 130 languages for Flash and 101 for Flash-Lite, with automatic language detection. Those lists establish supported languages, not equal quality across them.
An independent example shows the dialogue feature in use. Developer Simon Willison reports generating a 78-second conversation between two fictional pelicans with Flash in roughly 20 seconds, at a cost of 2.74 cents. That extrapolates to about $1.26 per hour, above Google’s current 81-cent output calculation. His post supplies neither billing records nor token counts, so the difference remains unresolved. It is a useful demonstration, not a verified all-in price or a test of voice replication.
Access also varies by feature. Google says Flash is available in Gemini Notebook and Flash-Lite in Google Vids, with Gemini Enterprise API access still forthcoming. Its launch footnote excludes AI Studio voice replication in Illinois, Texas, the EEA, the UK, Switzerland and India. That restriction expressly names AI Studio; the published material does not settle direct voice-replication API access in those places.
For copying a voice, Google requires more than the short sample. Its replication documentation describes a separate consent recording and a check that the two recordings contain the same adult speaker:
- Provide 10 to 30 seconds of clean, natural reference speech from the adult whose voice will be replicated.
- Provide a separate recording of that speaker reciting the exact consent statement in one of 30 supported locales. The English statement is: “I am the owner of this voice and I consent to Google using this voice to create a synthetic voice model.”
- Submit both recordings for the same-speaker check that Google says must pass before it creates the voice profile.
That is meaningful friction: under the documented procedure, an ordinary recording of someone speaking is insufficient by itself. But the statement authorizes creation of a synthetic voice model; it does not establish approval for every script later generated. And because Google accepts uploaded recordings, the workflow alone does not establish that the speaker is present or that the consent audio is authentic.
One developer’s forum report describes successful self-replication using a fresh 17-second Japanese reference and a separate 12-second consent recording. Pairing the same consent clip with an older reference produced a server error. The account includes no published audio or account records, and the error does not establish whether speaker matching, recording quality or another problem caused the failure. It reports no attempt using a different speaker or synthetic consent.
Google’s audio model card describes internal evaluations and red teaming, but its public materials provide no replication-specific false-acceptance rates, false-rejection rates or results for synthetic-consent attacks. That does not prove a bypass or rule out undisclosed safeguards. Our read is that the launch makes custom speech substantially easier to produce, while recorded consent remains a documented hurdle whose strength against deliberate impersonation has yet to be publicly demonstrated.