Google DeepMind introduces Gemini 3.8 Flash TTS for expressive audio generation

The new models allow for granular control over vocal delivery, including accent and emotional tone. According to Google, these models offer significant improvements in use cases such as long-form content and dual-speaker screenplay control compared to the previous Gemini 3.1 Flash TTS. Gemini 3.8 Flash TTS achieved high scores on Hume AI’s Voice Design Benchmark and lead in accent modeling.

In blind human preference evaluations on Voice Arena, both models secured top positions in key global languages, including Japanese. The models support over 100 languages and are available through Google AI Studio and the Gemini API.

To ensure responsible use, Google has implemented safety features, including consent verification for voice replication. All audio clips generated by Gemini Audio models are watermarked with SynthID, an imperceptible watermark designed to make AI-generated speech detectable.

PLUS ULTRAby Amenoyomi

The transition from static presets to a prompt-based system changes how vocal identities are created. Instead of selecting from a limited set of 30 original voices, users can now use natural language prompts to generate bespoke vocal personas from scratch. This allows for the creation of an infinite library of voices, ranging from a high-energy DJ from Melbourne to a monotone robot or a Japanese dragon.

Beyond the initial voice design, the models provide granular control over the performance of each line. Creators can direct specific accents and emotional tones to ensure the delivery matches the intended scene. This includes the ability to manage natural turn-taking in dialogue, enabling the transformation of scripts into fully performed scenes through a dual-speaker screenplay editor.