Gemini API: Voice Remixing Arrives for Gemini 3.8 TTS

Gemini CLINov 4, 2026View original changelog

Google added voice remixing to Gemini 3.8 TTS in the Gemini API. Developers can take a prebuilt, library or previously designed voice and change it with plain-language instructions such as age, pitch, energy, timbre or accent. Each remix produces a new, independent stored voice with a voice_... ID and an audio preview. The ID works with both Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS.

Key Takeaways

Voice remixing lets developers modify any existing Gemini 3.8 TTS voice with plain-language instructions and save the outcome as a new, independent voice.
Remixes start from a prebuilt voice, a voice library ID or a stored voice, so edits can be chained over several rounds, with Google advising one or two trait changes per round.
Each remix returns a voice_... ID and a 24 kHz WAV preview, which lets developers audition changes before using the voice in production.
The new voice works with both Gemini 3.8 Flash TTS and Flash-Lite TTS, so the cheaper, high-volume tier can reuse a voice created on the flagship model.
The feature needs a paid Tier 1 or higher project, allows 200 stored voices per project, and keeps voices for one year from last use.
Remixing is for lasting voice traits, while turn-level delivery cues still go in speech_metadata.style, and the base voice is never altered.

What Changed

Google added voice remixing to the Gemini 3.8 text-to-speech models in the Gemini API on November 4, 2026. The feature lets developers modify an existing voice using natural-language instructions and save the result as a brand-new voice. It sits alongside the prebuilt studio voices, the extended voice library, voice design and voice replication that already shipped with Gemini 3.8 TTS. Remixing can also be done interactively in Google AI Studio.

How Voice Remixing Works

A remix starts with a base voice and a short description of the changes. Developers call POST /v1beta/voices with the voice type set to prompted, store set to true, a base_voice and an input string. The base voice can be one of the 30 prebuilt studio voices (for example Puck, Kore or Zephyr), any catalog voice ID returned by the voice listing call, or any stored prompted voice that the project already owns.

The input describes relative changes rather than a whole persona. Google's own example takes the prebuilt Puck voice and asks for a voice that sounds "slightly older and more grounded, with a warm British accent". The model applies the edits and returns a new voice_... ID along with a 24 kHz WAV sample so the result can be auditioned before use. The original base voice is never modified, and deleting a base voice does not affect voices remixed from it.

Using and Chaining Remixed Voices

The new ID is passed in the speech configuration when synthesizing speech, and it works with both gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts. A remixed voice can also serve as the base_voice of another remix, so developers can refine a voice step by step. Google recommends changing only one or two traits per step and listening to each preview before the next one. By default a remix keeps the base voice's language and regional accent unless the instructions ask for a change.

Remixing is meant for lasting voice traits. Turn-level delivery cues, such as a whispered line or a rushed sentence, still belong in the speech_metadata.style field.

Limits and Management

Voice remixing requires a paid tier project (Tier 1 or higher).
Each project can hold up to 200 stateful voices, shared across designed, remixed and replicated voices.
Stored voices are kept for one year from last use, and every synthesis or use as a base voice resets that window.
Voices can be listed by type, which shows each remixed voice's base voice and instruction text for lineage tracking, retrieved to replay a preview, or deleted with DELETE /v1beta/voices/{id}.

Why It Matters

Before remixing, getting a slightly different voice meant designing one from scratch or cloning a recording. Remixing gives teams a faster way to tune a consistent brand or character voice, because small adjustments can be made, previewed and saved without starting over. Google's documentation does not state remixing-specific pricing or provenance rules, so developers should check the pricing page and the voice documentation for those details.