Design and replicate voices with Gemini 3.8 Flash TTS

Author: Google for Developers · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Design and replicate voices with Gemini 3.8 Flash TTS · Channel: Google for Developers · 14:04 · Watch on YouTube

Topics: Software engineering with AI, AI developer tools, AI content generation, AI models

Summary

The video demonstrates Gemini 3.8 Flash TTS features for designing custom voices from text prompts and replicating voices from audio samples, including a voice replicated from 20 seconds of audio. It shows how generated voices can speak different languages while retaining characteristics such as their designed accent, and how to use the Gemini API to manage custom and pre-built voices. The presenter also builds a tool for generating multi-speaker podcasts, with research, scripted dialogue, and assigned voice styles. The video closes by distinguishing Flash TTS from Flash-Lite TTS and highlighting partner integrations.

Key points

Google AI Studio can replicate a voice from a short audio sample or generate one from a text description.
Voice Design can generate multiple voice samples, along with sample text selected to reflect the designed voice.
A designed accent can carry over when the voice speaks another language.
Generated and replicated voices can be listed and reused through the Gemini API.
The presenter recommends installing Gemini API developer skills for current API and prompting guidance.
The extended voice library can be filtered by characteristics such as gender, keywords, or language.
Gemini-TTS supports multi-speaker dialogue, including back-channeling and distinct speaker styles.
The demo creates a researched podcast about Dublin with British English and Irish voices.
Gemini 3.8 Flash TTS is presented as expressive with broad language support; Flash-Lite TTS is positioned for high-volume, cost-efficient projects.
The video mentions integrations with Agora, LiveKit, Pipecat, and Vercel.

Chapters

Start	Chapter	Summary
00:03	Voice design and replication in AI Studio	The presenter introduces voice replication and creates a high-energy Argentine Spanish podcast host voice from a text prompt.
03:24	Accent behavior and API setup	A generated voice speaks German with a Spanish accent, then the presenter sets up the Gemini API tools and lists account voices.
07:23	Voice library and multi-speaker dialogue	The demo explores custom and pre-built voices before creating a tool for generating a two-host podcast.
10:48	Generating a Dublin history podcast	The tool researches Dublin, assigns British and Irish voices to the hosts, and generates a sample conversation.
12:57	TTS model options and integrations	The presenter compares Flash TTS and Flash-Lite TTS, points to benchmarks, and names partner integrations.

Quotes

from just 20 seconds of audio. — 00:35

High energy, mid-twenties female podcast host with a strong local Buenos Aires Argentine Spanish accent hosting a radio show in Spanish. — 01:09

It will actually have that accent in different languages as well. — 03:55

Another really cool feature of Gemini-TTS is the new multi-speaker dialog feature. — 08:30

the most expressive, most languages supported. — 12:57

Concepts and entities

Gemini 3.8 Flash TTS, Gemini 3.8 Flash-Lite TTS, Voice Design, voice replication, Google AI Studio, Gemini API, Antigravity IDE, Google, Agora, LiveKit, Pipecat, Vercel

Full transcript (original language)

00:03 Welcome back. Today we’re looking at something that completely shifts what’s possible with synthetic speech. Text-to-speech used to mean picking from a handful of preset voices that sounded decent, but never quite personal. That changes today. We’re diving into the latest TTS capabilities in Gemini, headlined by two massive upgrades. Voice design and voice replication. -Yes, exactly. So that was my voice that I just

00:35 replicated here in Google AI Studio from just 20 seconds of audio. So now Gemini-TTS can do all the heavy lifting for me. No of course, just kidding. I’m still here and I’m still here to talk you through the latest features. So what’s really cool here is we can replicate our voice right within Google AI Studio. We can also simply generate a new voice using Voice Design. And so we can simply

01:09 describe the voice that we’re looking for. And so for example, I’m not great at Spanish, but I just came back from a fun conference in Argentina. So Argentine podcast host. We’ll do that. And let’s design the following voice. High energy, mid-twenties female podcast host with a strong local Buenos Aires Argentine Spanish accent hosting a radio show in Spanish.

01:42 All right, there we go. And now we can also click and proof my prompt. And so now this will be a bit more detailed. And then we can click Generate Voice. And so now we will get three samples that we can choose from. So the model itself will both generate the voice as well as some sample text that it, decides itself to best reflect the new voice that we’ve designed. Here we are. So in our

02:16 case, the first voice generation failed. So we’re just seeing two options here. Let’s have a listen. - ¡Muy pero muy buenos días a todos, che! Ya estamos acá arrancando con todo esta mañana de locos que pinta tremenda. -Yeah that is definitely high energy. Let’s have a listen at the other one. - Che, ¡qué día, por favor! Me levanté y me cayó un rayo, ¿viste?, de la luz así por la ventana que dije: “¡Nah, esto es para la semana que viene, tengo a full!” -All right. That is definitely also very high energy. And speaking very quickly. Unfortunately, I don’t understand

02:47 what it’s saying. Gotta translate it later. But let’s choose this one for now. And so now we can what it’s saying. Gotta translate it later. But let’s choose this one for now. And so now we can use this voice. And we could also of course use this in any language whatsoever. So for example we could also have this voice speak German. Now this is a famous German poem. And now obviously this voice was created specifically for an Argentine Spanish accent. So, this might lead to a different result, but let’s have a listen.

03:24 -Frühling less sein blaues BandWeider flatten durch die Lufte Süße, wohlbekannte DüfteStreifen ahnungsvoll dan Land. Veilchen traumen schon, Wollen balde kommen. Horch, von fern ein leiser Harfenton Frühling ja du bist’s! Dich hab ’ich vernommen! -So you can hear. Now it’s actually speaking German with sort of a Spanish accent. Um, so that is something to keep in mind when you are designing a voice with a specific accent. It will

03:55 actually have that accent in different languages as well. So now that we have created some voices that we can use, we can reuse these in the API as well. So let’s switch over to our Antigravity IDE. So as a first step when working with the Gemini API. I recommend that you install the Gemini API dev skills, so you can just copy it from the docs here, and you can put it into your favorite

04:31 coding agent and simply install it globally, for example here. And so now we have our Gemini API dev skills installed, and then also know how to use the Gemini-TTS. It’s kind of the newest best practice instructions for the API, for the voices API, as well as for the prompting best practices. So what we can do here now is that we can again simply with voice instructions, create a simple Python script to list

05:05 all the voices that are that were created on my account. That includes both prompted voices as well as replicated voices. There we go. Let’s send this off. Okay, great. So now we have our script set up here. So we’ll accept these changes. We will use the latest Google Gen AI Python client. And then we will need to set up a .env file. So what we can do is we can simply copy

05:40 the .env.example into a .env file. And then in a .env file we can simply go back to our AI Studio. We can click on our API key here. And we can copy our API key that we want to use. Okay great. So I’ve set up my API key. And now we can have a look at our script here. So we’re just importing everything. We are

06:15 then using our API key, and we’re listing all our voices. So we can see here this is our Gen AI client. And then we have just the voices and we list them here. So first of all what we need to do is we need to um yeah we got our API key. And then what we’ll do is we will create a new let’s just say use UV, a new kind of virtual Python environment. And so now we’re

06:49 just updating our, README here. Okay. So accept this. So first of all. Okay. It’ll use our dependencies. So let’s just do uv. Run voices.

07:23 There we go and we can see I’ve been busy. I’ve got some prompted voices here. 22 of them, and then some replicated voice voices as well. And so we can see here we got our Argentine podcast host, we got Thor’s, studio voice, and, uh, we got a bunch of other tests that we’ve done here. We can also say, “create

07:57 a new script to list all the pre-built voices that are available on the Gemini API in the extended voice library.” Okay, great. So we can just accept these changes, and now we can see kind of the pre-built voices. We can filter for female voices or for keywords, or for example, for Japanese voices. Let’s

08:30 see. So one thing that’s really, really cool about Gemini-TTS, it’s really good across many different languages. And Japanese is also one of these languages that it’s really, really good at. And so let’s do the following. Another really cool feature of Gemini-TTS is the new multi-speaker dialog feature. So we can actually create a

09:05 multi-feature dialog, a multi-speaker dialog with back channeling and all the cool stuff. So let’s do this. Create a new script of a multi-speaker podcast generation tool. So we want to give the tool just a topic. Then it should use Gemini 3.8 Flash to briefly research the topic. Write a podcast script for two podcast hosts.

09:39 And we want one of these hosts to be a British English female and the other one to be an Irish male. All right. So now it’ll create a new script. Uh, hopefully it’ll search through all the pre-built voices for us to to find those specific voices that we’re looking for. So, because we have the coding agent skill, our coding agent can actually use the API key to

10:12 list the voices Itself and kind of complete all the tasks. Then in the end, we have a script ready that can create a podcast for us. Okay, fantastic. So our new podcast generation script is done. So we can see here, generate a podcast on any topic. We can also preview the script first if we want. We can specify certain voices.

10:48 So yeah you can see I have a custom Irish voice there. Already custom dialog length as well. So let’s do exactly this. We want a podcast with these voices about the history of Dublin.

11:19 Okay, so first of all, we’re using Gemini 3.8. Flash just like we instructed it to do the research. From Viking Longships to Silicon Docks. Okay. Nice. And then you can see here now that we are scripting kind of across, different personas. And then we are assigning a specific style to each section here. And now

11:53 our podcast is generated. And so let’s have a listen. -Liam. Now Charlotte most folks land in Dublin expecting cozy pubs. Fiddles playing in the corner and a fresh pint of stout, which is grand, of course. But underneath those cobbles is over a thousand years of sheer chaotic survival. -Quite literally chaotic. Even the name itself is an identity crisis, isn’t it? You have Dublin from the Old Irish, Dubh Linn, meaning Blackpool, and then

12:24 the modern Irish name Baile Átha Cliath. Right off the bat. Two different settlements competing for the map. -Spot on. -And you could see this is kind of three minutes of a multi-speaker dialog here. And that was generated pretty quickly which is really really cool. So Charlotte here British female and then Liam Irish male. We might need to list the voices to find a better British accent. So this is the gist.

12:57 This is the new Gemini 3.8 text-to-speech. You have the 3.8 Flash TTS version, which is the most expressive, most languages supported. And then we have Gemini 3.8 Flash-Lite TTS, really, for your high volume, cost efficient projects where you’re generating a lot of speech a little bit less language is supported, but still really, really strong in the benchmarks. So, what’s also really

13:31 cool is that this is integrated with many really cool partner integrations. So you can read the blog posts, you can find kind of all the benchmarks there. But you can also see that we’ve worked with, really cool partners like Agora, LiveKit, Pipecat, and Vercel to bring Gemini-TTS to their frameworks to really allow you to adopt these awesome models really quickly. So thanks so much for tuning in and I can’t wait to hear what you will

14:04 build with Gemini-TTS. Cheers!