Voice and audio
Gemini 3.8 TTS voice remixing: new NPC voices from a prompt
Google has added voice remixing to Gemini 3.8 TTS. You describe a change to an existing voice, such as older, lower or with a British accent, and the API stores the result as a new reusable voice.
On 4 November 2026 Google added voice remixing to the Gemini API’s text-to-speech models. It lets a developer take a voice that already exists and change it with a plain-language instruction, then keep the result as a new voice of its own.
The feature builds on Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, which Google launched on 23 September with voice design, the ability to create a voice from a written description.
How voice remixing works
You send a request to the new POST /v1beta/voices endpoint with a base_voice and an instruction in input. Google’s documentation gives the example of taking the prebuilt “Puck” voice and asking for something “slightly older and more grounded, with a warm British accent and measured audiobook narration pacing”.
The base can be any of:
- one of the 30 prebuilt studio voices;
- a voice from the Extended Voice Library;
- a voice you designed earlier, or one you have already remixed.
The API returns a new voice_... ID and a short 24 kHz WAV preview. The remix never changes the original voice, and deleting the original later does not affect the remixed copy. You can also remix interactively in Google AI Studio.
What it can do
- Change specific traits. Google lists age, pitch, energy, timbre and accent as things you can shift.
- Refine in steps. Because each remix is itself a stored voice, you can chain them: keep the same speaker and accent, but make the delivery calmer and a little lower.
- Use the voice on either model. A remixed voice works with both Gemini 3.8 Flash TTS and the cheaper Flash-Lite model.
- Track lineage. Stored voices record their base voice and the instruction that produced them.
Google’s September launch post says voice design works across more than 100 languages and dialects and that the voice library holds more than 2,000 voices. We have not tested how well a remix holds up in languages other than English.
Limits
The documentation describes the mechanism, not the quality. It gives no measure of how closely a remix keeps the original speaker’s identity, and the only listening evidence is the preview you generate yourself. Google’s launch post says voice replication comes with built-in consent verification; the remixing page does not say whether the same check applies to remixes of replicated voices, so check before building on that.
Our take
A game with a hundred named characters has an awkward voice problem: you want each to sound distinct, but you cannot cast a hundred actors, and a model that gives every character the same default voice breaks the illusion quickly. Remixing is a practical answer. A studio can design one core voice for a town, then derive the baker, the sibling and the older version of the same character from it, and store each as a stable ID.
The stable ID matters more than it sounds. An NPC that is meant to remember the player also has to sound like the same person on the next visit, which is the consistency problem we describe in what makes an NPC intelligent. Voice is only one layer of that: a stored voice does nothing for what the character knows, which is why memory in AI games still has to be solved separately.