Google introduced Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, two speech synthesis models designed to create original voices, reproduce authorized vocal profiles, and direct a performance with natural-language instructions. The release, announced on September 23, is already available in Google AI Studio and the Gemini API, as well as through integrations in Gemini Enterprise, Gemini Notebook, and Google Vids.

Two models for creation and scale

Flash TTS is the option with greater expressive capability. Google is targeting it at voice studios, video game characters, audiobooks, podcasts, and conversational agents that need to control tone, pacing, and nuance. Flash-Lite TTS is aimed at high-volume scenarios where cost and speed matter more, such as dubbing, mass content generation, and voice services.

The demonstration published by Google DeepMind shows two paths to creation. The first allows a completely new voice to be designed through a description: apparent age, timbre, accent, energy, or performance style. The second reproduces a consistent vocal profile from a 30-second sample, provided it is the user's voice or they have the right to use it.

The system also allows each line to be directed with instructions about emotion, pauses, speed, conversational sounds, and dialect changes. In scenes with two speakers, the creator can define distinct voices and adjust how they interact. This brings speech synthesis closer to an artistic direction workflow, rather than simply converting text into audio.

Controls against the risk of impersonation

Voice cloning expands legitimate uses—for example, maintaining an authorized narrator's voice across a series—but also increases the risk of fraud and impersonation. Google says the feature incorporates consent verification for reproduced profiles. Generated audio includes SynthID, the company's watermark for synthetic content, and C2PA credentials to preserve verifiable information about its provenance.

These controls do not eliminate abuse on their own. SynthID requires tools capable of detecting the watermark, while C2PA depends on platforms retaining the metadata and presenting it to the public. The model card indicates that Google conducted human evaluations, automated tests, and red-teaming exercises, but that page does not publish specific false acceptance rates for consent verification or detailed results on unauthorized cloning.

Performance and availability

Google says Gemini 3.8 Flash TTS ranked first overall in Hume AI's RW-Voice-EQ benchmark, with 71.4 points, and scored 60.8 in accent modeling. Hume describes RW-Voice-EQ as a multidimensional evaluation of voice systems under real-world conditions. Even so, an aggregate ranking does not by itself determine quality for every language, accent, or production use case; teams will need to test naturalness, consistency, and latency with their own data.

The release extends competition in generative audio toward a specific combination: voice creation through instructions, authorized reproduction, detailed direction, and deployment at scale. For developers and producers, the advance narrows the gap between a prototype and a production tool. For platforms and organizations, it also increases the need to preserve provenance credentials and verify authorization before publishing synthetic voices.

The Gemini 3.8 Audio model card groups Flash TTS and Flash-Lite TTS with the Live models in the same family. Google says these variants do not represent a material increase in frontier capabilities compared with Gemini 3.7 Flash. That assessment refers to its frontier safety framework; it should not be interpreted as a complete measurement of the social risk of voice cloning.