Gemini TTS voice remixing rewrites a voice from a prompt

Google's Gemini 3.8 Flash TTS can now remix a voice. On November 4, Google released Voice remixing for Gemini 3.8 Flash TTS in the Gemini API and Google AI Studio, a feature that takes a voice a developer already has and rewrites its timbre, pitch, pace, and accent from a written instruction, then returns a new voice to keep.

The release closes a gap that Google itself flagged. When the company put Gemini 3.8 Flash TTS and its lighter sibling into general availability on September 22, it listed voice remixing as "coming soon" alongside the voice library and replication. Six weeks later it shipped. That matters to anyone building narration, audiobooks, voice agents, or localized dialogue, because the older way to get a voice you half liked was to rebuild it from scratch and hope.

What Voice remixing does

Voice remixing modifies an existing voice using natural-language instructions to create a new, independent stored voice, according to the Voice remixing guide. It is not cloning and not designing from nothing. The developer picks a base voice that already works and describes the relative changes they want, such as shifting the age, pitch, energy, timbre, or accent. The model applies those changes to the base voice's acoustic profile and stores the result as a separate resource.

The guide is explicit about what does not happen: remixing always creates a new, standalone voice, and it never modifies or overwrites the original base voice. Deleting the base voice later does not affect any voice remixed from it.

How it works in the Gemini API

A remix is one call to the Voices endpoint. The developer calls POST /v1beta/voices with gemini-3.8-flash-tts, sets type="prompted" and store=True, and points prompted.base_voice at a voice they already have. The model returns a new voice_... ID plus a 24kHz WAV sample_audio preview, which can be saved and replayed.

That ID then works for speech synthesis on both Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, so one remixed voice serves both the expressive model and the cheaper, higher-throughput one. The release notes describe the same two paths: the API call or the interactive workflow in Google AI Studio.

What you can change, and how you hear it

Google for Developers' own announcement post lists the adjustable traits as timbre, pitch, pace, and accent, and gives "soften the delivery" as an example instruction. The company's launch blog adds "add subtle Southern US accent" and a stylized voice for a fire-breathing dragon among the prompt examples. The guide frames the instruction as relative: because the model anchors on the base voice, a good prompt names only the traits to change rather than re-describing the whole persona, and a remix prompt preserves the base voice's language and regional accent unless a new one is stated.

The post also describes the studio flow. In Google AI Studio, a developer picks a base voice from over 2,000 preset library voices or a custom designed voice, enters a prompt, and auditions three generated takes before saving a favorite.

Where the base voice comes from

The guide allows three starting points. A developer can pass any of the 30 prebuilt studio voice names, such as Puck, Kore, or Zephyr, or any catalog voice ID returned by the Voices list call. They can start from a Voice design persona they created earlier, or from a voice they remixed in a previous step. Because every remixed voice is itself a stored resource, remixes can be chained, refining one or two traits at a time from a designed or previously remixed voice. Remixed voices keep their lineage in the record: a stored voice carries both its base_voice and the input prompt, so a project can trace where each one came from.

What it costs and the limits to plan around

Voice remixing requires a paid Gemini API project at Tier 1 or higher, per the guide. A project can hold up to 200 stateful stored voices under store=True, a pool shared across designed, remixed, and stored replicated voices. Stored voices carry a one-year retention window from last use, and the clock resets each time a developer synthesizes speech with the voice or references it as a base voice in a new remix. The 30 prebuilt studio names that ship with the API are a small set next to the library: the guide's supported voices list cites 30 prebuilt studio names, while the Extended Voice Library catalog holds the rest.

How remixing differs from design and replication

Gemini's voice tooling now has three parts, and Google's September 22, 2026 release notes grouped all of them under the TTS launch. Voice design creates a persona from a text prompt. Voice replication recreates a speaker from a 30-second audio sample, and Google's launch blog says replication ships with consent verification, SynthID watermarking, and C2PA credentials. Voice remixing sits between them, adjusting a voice that already exists without a real-person recording.

The three differ in where they are offered, too. AlphaSignal reports that Google AI Studio does not offer replication in Illinois, Texas, the European Economic Area, the United Kingdom, Switzerland, or India, a restriction that applies to replication from a real-person sample rather than to preset or designed voices. The same AlphaSignal report says Gemini 3.8 Flash TTS ranks first overall on the third-party Hume AI Voice Design Benchmark with a score of 71.4, and first for accent modeling at 60.8. SLOP TV has not verified that benchmark independently.

The September launch and the November 17 shutdowns

The TTS models and the Voices endpoint reached general availability on September 22, 2026, according to the release notes, alongside voice design, voice replication, and the Extended Voice Library. Voice remixing arrived on November 4 as the last piece of that set.

It also arrives just before a migration deadline. Google will shut down three older preview endpoints on November 17, 2026: gemini-2.5-flash-preview-tts, gemini-2.5-pro-preview-tts, and gemini-3.1-flash-tts-preview. Google names Gemini 3.8 Flash TTS and Flash-Lite TTS as the replacements for all three. The pair shares one release date, September 22, 2026, on the deprecation page.

What it means for creators

For a studio that has a voice close to what a script needs, remixing is the cheaper move: adjust the base voice instead of commissioning a new persona or re-recording. The audition step matters for anyone who has to approve a voice, because it puts three takes side by side before a decision rather than after a render.

Two limits are worth weighing first. The paid-tier requirement and the 200-voice ceiling both shape how a project plans its library, and the one-year retention window means a voice left idle needs a fresh synthesis or remix to keep its place. The geography rule AlphaSignal reports is about replication, not remixing, so the practical check for a remix workflow is which base voices a project can already use.

SLOP TV has not tested Voice remixing. Every figure here comes from Google's own documentation, its developer posts, or the AlphaSignal report named in the text.

Voice remixing is available now in the Gemini API through POST /v1beta/voices with gemini-3.8-flash-tts, and interactively under "Remix this voice" in Google AI Studio.