Google Gemini 3.1 Flash Live debuts

This is less “TTS API” and more Google turning voice into a first-class app interface for agents. The model matters because it combines conversational quality, latency, and tool execution in one stack instead of treating speech as a bolt-on.

• –Available in preview through the Gemini Live API, with Search Live and Gemini Live already powered by it in 200+ countries
• –Google says it improves tone understanding, long-horizon conversation retention, and function-calling reliability in noisy, real-world settings
• –Benchmarks cited by Google put it ahead on ComplexFuncBench Audio and Audio MultiChallenge, which is the kind of signal developers care about for voice agents
• –SynthID watermarking on generated audio is a practical safety move, especially if this gets used in dubbing, assistants, or content workflows
• –The release strengthens Gemini’s position against other “voice layer” offerings by making natural conversation, multilingual support, and agentic execution part of the same model