On 22 September 2026 Google made two new Gemini text to speech models generally available and opened a Voices endpoint in the Gemini API. It adds voice design, voice replication and a larger library of voices, making spoken audio a normal output for AI apps.
Key facts
- `gemini-3.8-flash-tts` and `gemini-3.8-flash-lite-tts` reached general availability on 22 September 2026.
- The Flash-Lite TTS model replaces `gemini-3.1-flash-tts-preview`.
- A new Voices endpoint (`/v1beta/voices`) launched in the Gemini API.
- It brings voice design, voice replication with consent verification and 150+ prebuilt and custom voices.
What happened
Text to speech, or TTS, turns written text into spoken audio. In its 22 September release notes, Google made two new TTS models generally available: Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS. The lighter model is meant to replace an earlier preview.
Google also released a Voices endpoint in the Gemini API. With it, developers can design custom voices from text prompts, replicate voices with consent verification and query more than 150 prebuilt and custom voices. A week earlier, on 15 September, Google had also made its Gemini 3.8 Live models generally available for real time audio conversations.
Together, these releases make voice a standard building block for apps built on Gemini. Possible uses range from narrated videos and podcasts to voice assistants and accessibility tools.
Why it matters
Voice is quickly becoming as common an AI output as text and images. Short videos, audio summaries, product explainers and podcasts can now be produced with AI voices in minutes.
But a script written for the eye does not sound good out loud. Long sentences, brackets and lists that work on a page become hard to follow when spoken. Writing for the ear means short sentences, one idea at a time, clear transitions and a defined tone. The better the script, the more natural any AI voice will sound.
Voice replication also raises obvious questions of consent and trust. Only replicate a voice when you have clear permission, and say when audio is AI generated.
How to try it
A practical way to start is with short social content, where voice overs are everywhere. Our free prompt Plan 7 social posts creates a week of posts with different formats and a call to action for each day. In the brackets, describe your business or project, your audience and the platform, for example “a small bakery, local families, Instagram Reels”.
When you have the plan, pick one post and ask the assistant: “Write a 30 second voice over script for this post. Short sentences, friendly tone, no lists, one clear call to action at the end.” Read it out loud yourself before using any AI voice: if you stumble, the AI voice will too.
Check the script for length (about 70 to 80 words for 30 seconds) and make sure every claim about your business is true.
Frequently asked questions
What did Google release for Gemini voices?
Two text to speech models, Gemini 3.8 Flash TTS and 3.8 Flash-Lite TTS, and a Voices endpoint in the Gemini API with voice design, voice replication with consent verification and 150+ prebuilt and custom voices.
What is voice replication?
It is the ability to create a synthetic voice that sounds like an existing one. It should only be used with clear permission from the person whose voice it is.
How do I write a good script for an AI voice?
Use short sentences, one idea at a time, natural transitions and a defined tone. Read it aloud before generating the audio.
More news
Nano Banana 2.1: Google’s new image model is here
An update to Nano Banana 2 with better quality, prompt adherence and text, at Flash-level speed and cost.
Gemini 4 Carbon: Google’s new model for coding and security
Announced two weeks after Argon, it starts with cyber defenders and reaches everyone later.
Gemini 4 Argon is generally available: what to know
Google's most capable model so far, built for complex coding, agents, math and science.