What Google shipped
On 4 November 2026 Google added voice remixing to the Gemini API changelog for Gemini 3.8 Flash TTS (text-to-speech). Voice remixing takes a voice you already have, applies changes you describe in plain English and saves the result as a new voice. Google for Developers announced the feature the same day. The announcement says it works in Google AI Studio and the Gemini API, and that you can start from more than 2,000 preset library voices or from a voice you designed yourself. TokenPost's report from 5 November lists the same controls: timbre, pitch, pace and accent.
Remixing sits on top of the TTS models Google released on 23 September. The Next Web's launch coverage described Flash TTS as the model for creative work like games and audiobooks, and Flash-Lite TTS as the cheaper option for high-volume dubbing and voice agents. At launch Google said remixing was coming soon, and the 4 November release delivered it.
How the Voices API works
Remixing uses the same endpoint as voice design, POST /v1beta/voices. According to the voice remixing docs, the request has these fields:
store: must betruefor prompted voices, so the voice is saved to your project and gets an ID.voice.model: defaults togemini-3.8-flash-tts.voice.type: set to"prompted".voice.prompted.base_voice: one of the 30 prebuilt voices (such as "Puck" or "Kore"), a catalogue voice ID fromGET /v1beta/voices, or a storedvoice_...ID your project owns.voice.prompted.input: the changes to apply, such as age, pitch, energy, timbre or accent.
The response returns a new voice_... ID and a sample_audio preview, a 24kHz WAV file. You can then use that ID with either gemini-3.8-flash-tts or gemini-3.8-flash-lite-tts. Remixing never changes the base voice. Deleting the base voice also leaves any voices remixed from it untouched. The docs say a remix keeps the base voice's language and regional accent unless your prompt asks for a different one.
Building a remixed voice
The docs' Python example uses the google-genai SDK and expects your key in the GEMINI_API_KEY environment variable. The block below follows the structure of the docs' example and keeps the docs' prompt (shortened here):
from google import genai
client = genai.Client()
remixed_voice = client.voices.create(
store=True,
voice={
"model": "gemini-3.8-flash-tts",
"type": "prompted",
"prompted": {
"base_voice": "Puck",
"input": "Make the voice sound slightly older and more grounded, ...",
},
},
)
print(remixed_voice.id)The docs then base64-decode sample_audio into a WAV file, so you can hear the preview before you commit to the voice. To generate speech, the remixing guide calls client.interactions.create with gemini-3.8-flash-tts, a user turn holding the text, and response_format set to audio. The new voice ID goes in generation_config under speech_config as {"voice": remixed_voice.id}. The returned audio is in output_audio.
To audit or clean up your voices, the docs show two calls:
response = client.voices.list(type_=["prompted"])
client.voices.delete(id="voice_abc123...")The list call returns each voice's base voice and the prompt that created it, so you can trace a chain of remixes back to the original preset.
Getting good results
You can chain remixes by passing a stored voice ID as the next base_voice. The docs' second example prompt keeps the speaker's identity and accent, then asks for a calmer, softer and slightly lower delivery. Google's guidance is to describe only the traits you want changed, and to change one or two traits per step while checking the preview each time. A remix that rewrites the whole persona drifts further from the voice you started with.
The voice design docs separate lasting traits from delivery. Age, timbre and accent belong in the voice prompt when you create the voice. Emotion and delivery that change from line to line, such as "whispered urgently", go in speech_metadata.style on each synthesis request. A game character with one stored voice can then sound calm in one scene and panicked in the next without needing a second voice.
Limits and costs
The remixing docs list these constraints:
- Paid tier only: remixing needs a paid Gemini API project (Tier 1 or higher).
- 200 stored voices per project: designed, remixed and replicated voices all share this cap.
- Expiry: a voice expires one year after its last use. Generating speech with the voice, or using it as a base voice, resets the clock.
The Gemini 3.8 Flash TTS model page lists an input limit of 8,192 tokens and an output limit of 16,384 tokens, which caps how much audio one request can return. Long scripts need splitting across several requests.
On the pricing page, Flash TTS costs US$0.50 per million input tokens and US$9.00 per million audio output tokens until 31 December 2026. Google puts that at about US$0.00225 per 10 seconds of audio. From 1 January 2027 both prices double, to US$1.00 and US$18.00. Flash-Lite TTS audio output costs US$6.00 per million tokens until the end of 2026 and US$12.00 from January. Synthesis on the free tier costs nothing, but remixing itself needs the paid tier. The pricing page doesn't list a separate charge for creating or storing voices, and Google hasn't said whether one exists. If you're budgeting a long audiobook or a busy voice agent, plan around the 2027 rates, since anything you launch now will run at them within weeks.
What it means for New Zealand builders
Accent control is the obvious draw for local products, but the docs don't confirm how well it works for us. The speech generation docs list English as one language entry with no regional variants. Māori isn't in the supported-languages table, although the models detect the input language automatically. Whether a prompt like "a light New Zealand accent" gives a convincing result is unknown. A cheap way to find out is to remix one preset, listen to the sample_audio preview and get Kiwi listeners to rate it before you build a product around it. Te reo Māori pronunciation in a remixed voice needs the same testing, and a fluent speaker should do it.
Remixing doesn't clone anyone's voice, but voice replication on the same platform does. The Next Web reported that Google puts a SynthID watermark on audio from its Gemini Audio models and requires a spoken consent recording before it clones a voice. If you prompt a remix toward a well-known speaker's sound, you're working outside anything the docs approve, so check Google's usage policies first.
Good first projects are audiobook narration with one consistent narrator, a cast of game characters chained from a single base voice, and a branded voice for a phone agent that you then run on Flash-Lite for cost. If you build one, share it on the ASAI projects page.
Sources
- Gemini API voice remixing documentation
- Gemini API changelog
- Gemini API voice design documentation
- Gemini 3.8 Flash TTS model page
- Gemini API pricing
- Gemini API speech generation documentation
- Google for Developers on X: Voice Remixing is now live in Gemini 3.8 Flash TTS
- TokenPost: Google Adds Prompt-Based Voice Remixing to Gemini 3.8 Flash Text-to-Speech
- The Next Web: Google's new Gemini TTS models can clone a voice from 30 seconds of audio
Build with people working on the same things
Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.