Google DeepMind has introduced Gemini 3.1 Flash TTS, a new text-to-speech model that aims to push expressive audio generation forward. The headline feature is a set of granular audio tags that give creators direct, precise control over how the model renders speech. Instead of relying on broad style presets, users can specify fine-grained vocal attributes to shape the output.
According to the announcement, these tags allow for explicit direction of AI speech, enabling more nuanced performances than typical TTS systems. The model is designed for use cases where emotional nuance, pacing, or emphasis matter—such as audiobooks, game dialogue, or interactive media. DeepMind frames this as the next step in making synthetic speech feel more natural and intentional.
The blog post does not provide technical details on the underlying architecture or benchmark comparisons. It focuses on the capability and the intended creative control. As the only source, it presents the model as a leap forward in expressive TTS, but independent evaluations are not yet available.