Tech • AI • Robotics • Game
Design and replicate voices with Gemini 3.8 Flash TTS
8/10TL;DR
Google has expanded Gemini TTS with voice replication, prompt-based voice design, multilingual output, reusable API access, and multi-speaker dialogue generation aimed at both expressive and cost-efficient speech applications.
KEY POINTS
Two major upgrades
Gemini TTS now adds voice replication and voice design, moving beyond fixed preset voices. A voice can be cloned from about 20 seconds of audio, while a new synthetic voice can be created from a written description such as age, energy, accent, and role.
Prompted voice creation
In Google AI Studio, users can describe a target persona and generate multiple sample voices to choose from. One example used a high-energy Buenos Aires Argentine Spanish radio-style host, and the system produced sample speech automatically to showcase the designed voice.
Accent carries across languages
Designed voices can be reused in languages other than the one they were created for. In testing, a voice built for Argentine Spanish read a German poem with a noticeable Spanish accent, highlighting that accent traits persist across multilingual output.
Voices available through the API
Custom voices created through prompting or replication can be reused programmatically through the Gemini API. A simple Python setup using the latest Google Gen AI client can list voices tied to an account, including both designed and replicated voices.
Large personal voice libraries are possible
Account-level management supports multiple saved voices. In one demonstration, the account contained 22 prompted voices plus several replicated voices, including named custom profiles and test voices, indicating support for broader experimentation and production workflows.
Pre-built extended voice library
Beyond custom voices, the platform also offers an extended library of pre-built voices. These can be filtered by attributes such as gender or language, including Japanese, which was highlighted as one of the stronger languages for the system.
Multi-speaker dialogue generation
A new multi-speaker mode can generate back-and-forth conversations with distinct personas and delivery styles. A sample workflow created a two-host audio program by first using Gemini 3.8 Flash to research a topic, then writing a script and assigning different voices to each host.
Example: history of Dublin
One generated segment featured a British English female host and an Irish male host discussing the history of Dublin, from Viking origins to the modern Silicon Docks era. The result was roughly three minutes of dialogue produced quickly, though accent selection still benefited from voice-library tuning.
Two TTS model tiers
The lineup includes Gemini 3.8 Flash TTS, positioned as the more expressive option with the broadest language support, and Gemini 3.8 Flash-Lite TTS, aimed at high-volume, cost-efficient use cases. The lighter model supports fewer languages but is still presented as strong on quality benchmarks.
Partner integrations
Adoption is being pushed through integrations with partners including Agora, LiveKit, Pipecat, and Vercel. These partnerships are designed to make the TTS models easier to deploy in real-time communications, voice agents, and application frameworks.
CONCLUSION
The latest Gemini TTS release pushes synthetic speech from generic presets toward customizable, multilingual, and programmable voice production. The main competitive question now is how reliably these tools can balance realism, accent control, and scalability in real-world products.
Ask a question