Google Launches Gemini 3.1 Flash Live: Real-Time AI Voice in 90+ Languages

Google DeepMind released Gemini 3.1 Flash Live on March 26, a new real-time multimodal voice model now available in preview through the Gemini Live API in Google AI Studio.

The model is Google's most advanced audio model to date, engineered to eliminate the latency stack that has plagued previous voice AI systems. Rather than chaining voice activity detection, transcription, language model generation, and text-to-speech sequentially, Gemini 3.1 Flash Live processes audio natively and collapses the entire pipeline into a single operation.

Key capabilities:

The model processes acoustic nuances in real time, recognizes pitch and pace, and performs reliably in noisy environments. It supports barge-in, allowing users to interrupt the AI mid-sentence — mimicking natural human conversation.

Gemini 3.1 Flash Live is fully multimodal: it accepts text, images, audio, and video inputs, and produces audio and text outputs. Video streams are processed as sequences of JPEG or PNG frames.

Developers can tune the model's reasoning depth using a thinkingLevel parameter with four levels — minimal, low, medium, and high — to optimize for latency versus problem-solving depth.

All audio outputs include Google's SynthID watermark, an imperceptible digital tag to help detect AI-generated content.

Global reach:

The model supports over 90 languages and has enabled the global rollout of Search Live to more than 200 countries and territories. This is Google's answer to the rapidly evolving voice AI sector, where competitors like OpenAI and ElevenLabs are pushing real-time interaction boundaries.

What this means for CIOs:

For enterprises evaluating voice AI in customer service, internal assistance, or multilingual support, Gemini 3.1 Flash Live represents a meaningful step forward. Low latency, 90+ languages, and noise robustness make it relevant for industrial and logistics-heavy operations.

API access is now available to developers via Google AI Studio.