Create and Deploy Custom Audio With Gemini 3.8 Flash TTS: Voice Design, Cloning, and the Two Clocks That Decide

Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS both let you invent a voice from a written description instead of picking one off a list, and both went generally available on the Gemini API on September 23, 2026. The headline people are repeating is the voice design. The part worth planning around is what happens to a designed voice afterwards, because the vendor gives you two ways to keep one and attaches a different lifetime to each. Choose the wrong one and the voice your product depends on expires in a week.

That is the gap in the launch coverage so far. The announcement posts explain that you can prompt a voice into existence, and a few of them mention cloning from a 30-second sample. None of them walk through the storage model, which is the thing that decides whether a speech pipeline is reproducible from a config file or has to be re-provisioned on a schedule. This piece covers the whole path — design, store, call, and pay — with the prices and the quotas as Google publishes them.

What shipped on September 23

Two models, one API schema. The identifiers are gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts, and Google's documentation is explicit that both take "the exact same API schema and prompting format" — moving between them is a change to one parameter, not a rewrite.

• Input — text only. Both models accept text and return audio; they are not multimodal and do not generate text back.

• Output — WAV, 24 kHz mono 16-bit signed PCM on the unary endpoint; headerless PCM (audio/l16) on the streaming endpoint; audio/mulaw and audio/alaw accepted with a configurable sample rate.

• Languages — 130 on Flash TTS, 101 on Flash-Lite TTS, auto-detected from the text rather than declared in the request.

• Voices — 30 curated studio voices listed in the docs (Kore and Puck among them), an Extended Voice Library of "hundreds" more queryable through the voices endpoint, plus the two creation paths below.

• Direction — line-by-line delivery control, two-speaker scenes staged from a single script, and non-verbal cues such as <laughs>, <sigh> and |mhm| written straight into the transcript.

The two tiers exist because the workloads diverged. Flash TTS is positioned for maximum acoustic fidelity and complex acting; Flash-Lite TTS is described in Google's own words as the "workhorse replacement" for gemini-3.1-flash-tts-preview, aimed at bulk dubbing, read-aloud features and voice-agent cascades. Both replace a single preview model that had been on the API since April 2026, so if you were on the preview, the migration is a tier decision rather than a version bump.

Designing a voice is a prompt, and that changes the review step

The creation surface is the genuinely new thing in this generation. A prompted voice is built from a natural-language description — role, accent, vocal character — and returns an identifier you reuse. In the API this is a voice-creation call with store: true and a prompted input; in Google's own examples the descriptions run from a high-energy Melbourne DJ to a Japanese dragon. Once created, the voice is referenced in a speech request by its identifier, and the per-request style instructions can then be minimal or empty, because the character is already baked into the voice.

The practical consequence is a workflow change, not just a feature. Voice casting used to be a licence negotiation or a studio booking, which meant the voice selection happened once, early, and survived review because changing it was expensive. When a voice is a paragraph of text, casting becomes a design token — something a product manager can ask to have re-rolled on Tuesday. Teams that put the description string in source control and treat the generated voice ID as a build artifact will get consistent output across environments. Teams that create voices by hand in AI Studio will not, and the failure will look like an inexplicable difference between staging and production audio.

The two clocks

This is the section the launch posts skip. Google lets you keep a designed or replicated voice in one of two states, and they expire at wildly different rates.

• Stored voices — created with storage enabled, they live in your project and are governed by a quota of 200 voices per project with a one-year time-to-live.

• Stateless keys — voice replication can return a client-managed key (the voicekey_… form) instead of a stored identifier, and that key has a seven-day lifetime.

Read together, that pair is a design instruction. A stored voice is a year-long commitment and a hard ceiling of 200 per project, which is generous for a product with a fixed cast and useless for anything that generates a new voice per customer. The stateless key is the opposite: no quota pressure, but a key you have to re-mint weekly, which means your application needs a refresh path and your infrastructure needs to store credentials that rotate. Neither is wrong. Picking without noticing which one you picked is how a voice agent goes silent on day eight.

Two more properties come attached to replication and are worth knowing before you promise anything to a legal team. Creating a replicated voice requires a verbal consent recording from the voice's owner that must match the reference speaker — a spoken check, not a checkbox. And the output carries provenance: every clip is watermarked with SynthID, and replicated voices additionally carry C2PA content credentials. The design path is deliberately the harder one to abuse, and both barriers are structural rather than policy statements.

One discrepancy worth flagging rather than resolving. Google's supported-models table lists voice design and voice replication as available on both 3.8 TTS models. Most launch coverage describes replication as part of the Flash TTS story specifically. If replication matters to your decision between the tiers, verify it against the docs for the tier you intend to ship rather than trusting either summary.

What an hour of audio costs

Google bills speech in tokens, not characters or seconds, and the conversion is published: audio is metered at 25 tokens per second of output, which is 90,000 tokens per hour. That makes the arithmetic unusually clean, and it is the number to put in a budget rather than the per-million rate.

• Flash TTS standard — $0.50 per million text tokens in, $9.00 per million audio tokens out through December 31, 2026, then $1.00 and $18.00. Batch and Flex are $0.25 and $4.50; Priority is $0.90 and $16.20.

• Flash-Lite TTS standard — $0.50 and $6.00 through the same date, then $1.00 and $12.00. Batch and Flex $0.25 and $3.00; Priority $0.90 and $10.80.

• In seconds — Flash TTS works out at roughly $0.81 per hour of generated audio during the promotional window and about $1.62 after it; Flash-Lite TTS is roughly $0.54 and then $1.08.

• Against the model it replaces — gemini-3.1-flash-tts-preview is priced at $1.00 and $20.00 per million, so the audio-output line fell 55 percent on Flash TTS and 70 percent on Flash-Lite TTS.

Do not plan against the promotional number. Google publishes both columns in the same table, which means the January rate is not a rumour to be discovered later — it is on the card today, and any per-thousand-minutes estimate built at $0.81 has to survive being doubled.

One independent number, and several that are not

Google's announcement leads with benchmark results, and they should be read for exactly what they are. The company says Flash TTS took first place on Hume AI's Voice Design Benchmark at 71.4, led accent modelling at 60.8, and that Flash TTS and Flash-Lite TTS ranked first and second on Hume's Overall Quality Index, with strong blind-preference placings on Voice Arena for Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic, Mexican Spanish and Hindi. All vendor-reported, none of the runs public.

There is a detail in that list worth noticing. Alan Cowen, credited as an author of Google's announcement, is the founder of Hume AI — the lab whose Voice Design Benchmark supplies the headline number. That does not make the figure wrong, and Hume's benchmark is a real one, but the two are not independent of each other, and a reader weighing the 71.4 should know that before repeating it as third-party validation.

The independently collected figure is on Artificial Analysis's Provider Voice Arena, captured for this article on September 24, 2026: Gemini 3.8 Flash TTS sits second at an Elo of 1,260 with a 17-point interval across 1,999 samples, behind Cartesia Sonic 3.6 at 1,273. Gemini 3.8 Flash-Lite TTS is sixth at 1,235. The model being replaced, Gemini 3.1 Flash TTS, is tenth at 1,199 across 3,430 samples. Blind-listener preference, not a lab's own eval — and the only quality number here that Google did not commission.

Where to run it, and where it is not yet

Both models are available today in the Gemini API and Google AI Studio, with no waitlist. The consumer and enterprise surfaces are staged rather than shipped: Flash TTS is coming to Gemini Notebook and Flash-Lite TTS to Google Vids, Gemini Enterprise access is described as coming soon, and voice remixing — adjusting timbre, pitch, pace and accent on an existing voice — is listed as a future feature rather than a present one. If your plan depends on remixing, it does not exist yet.

On our own side, honesty first: the Gemini 3.8 TTS endpoints are not on OrcaRouter's routing table. The generation they replace is — google/gemini-3.1-flash-tts-preview is in the catalogue at the same $1.00 and $20.00 per million tokens Google publishes, because provider list price is passed through with no markup, and it answers on the same /v1/audio/speech endpoint your OpenAI-compatible SDK already targets. That matters more than it sounds for a speech workload, because the thing you are actually buying is a stable endpoint your application can call while the models behind it change. Your code holds a model string; the catalogue holds the routing. When Google's promotional window closes on December 31 and Flash TTS doubles, a routed model's price moves the same day rather than at the next contract renewal — and a model that goes down fails over instead of taking your narration queue with it.

If you are evaluating this generation before committing, the cheap experiment is a two-tier one. Run the same script through Flash TTS and Flash-Lite TTS, because the schema is identical and the difference is a parameter — then decide with your own audio rather than with a benchmark chart, and check on Artificial Analysis whether the quality gap survives its error bars before you pay the premium. For a large-format narration project, the premium is real money; for a support bot reading a phone number aloud, it is not obviously buying you anything a listener can hear.