Latest in AI
Gemini 3.8 TTS Voice Remixing: API Access, Stored Voices and Pricing
Google now lets developers turn an existing synthetic voice into a reusable variant, with paid-tier access, storage limits and separate speech-generation costs.
Google released voice remixing for Gemini 3.8 TTS on November 4. Developers can describe changes to an existing prebuilt or designed voice, save the result as a new voice and reuse it for speech generation. The API release notes confirm availability through the Voices endpoint and Google AI Studio.
This is the arrival of a feature previously listed as coming soon, rather than a new TTS model launch. For a team producing narrated lessons or product audio, it creates a way to refine a voice without starting each variation from an empty prompt.
How a voice remix works
Google's voice remixing guide uses POST /v1beta/voices with gemini-3.8-flash-tts, a base voice and instructions describing the desired changes. The result has its own stored voice_... identifier and an audio sample. Gemini 3.8 Flash TTS and Flash-Lite TTS can both synthesize speech with that identifier.
Supported starting points include prebuilt voices, Extended Voice Library voices and stored prompted voices owned by the project. The guide does not list replicated voices as supported remix bases. The source voice stays unchanged, and deleting it does not delete its remixed descendants.
That independence is useful for versioning. A product team could keep its approved narrator while evaluating a variant for a different audience. It should still maintain a record of which voice ID belongs to each published asset, so later experiments do not accidentally change an established production choice.
Access requirements and storage limits
Remixing requires Paid Tier 1 or higher. Designed, remixed and stored replicated voices share a limit of 200 stored voices per project. A voice expires after one year without use; synthesis or reuse as a remix base resets that period.
Google distinguishes durable voice characteristics from temporary performance instructions. Use remixing for the former and speech_metadata.style for the latter. Language and accent stay with the base voice unless explicitly changed.
For a large audio catalogue, this makes voice inventory part of operations. Keep a mapping between voice IDs and their intended uses, review unused experiments, and test how the application handles an unavailable ID. A successful sample does not establish that the voice will pronounce every product name or maintain the desired delivery across a long script.
Gemini 3.8 TTS synthesis pricing
The published pricing lists these Standard paid-tier rates in US dollars per million tokens through December 31, 2026:
| Model | Text input | Audio output |
|---|---|---|
| Gemini 3.8 Flash TTS | $0.50 | $9.00 |
| Gemini 3.8 Flash-Lite TTS | $0.50 | $6.00 |
On January 1, 2027, the listed rates rise to $1 for text input and $18 or $12 for audio output, respectively. Google counts audio at 25 tokens per second.
These are speech-synthesis rates. The page does not specify a separate remix-creation charge, so they should not be presented as a complete price for creating and storing a voice.
Compare the two synthesis models using the same approved voice and script. Listen for pronunciation, pacing and continuity, then account for the retries needed to obtain usable audio. That gives a more practical cost comparison than choosing from token rates alone.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help technical founders turn AI product releases into clear explanations and launch stories that reach developers and customers.