On March 26, Google released Gemini 3.1 Flash Live.
If you think this is just a faster, cheaper voice synthesis API, it's worth a second look. This model makes a fundamental change in technical architecture.
How Traditional Voice AI Works
The old AI voice interaction was essentially a three-stage pipeline:
- Speech Recognition (ASR): Converts your audio to text
- Language Model Inference: Text goes into the LLM, generating a text response
- Speech Synthesis (TTS): Converts the text back into voice output
Each step introduces latency, creating that feeling of waiting for a reply after speaking. Moreover, each conversion loses information — your tone, pauses, and emotions are lost in the text stage.
What Gemini 3.1 Flash Live Does
It compresses this three-stage pipeline into a single native audio-to-audio model.
Without going through text as an intermediate state, it goes directly from your voice input to voice output.
Technical specifications:
- Audio input: 16-bit PCM, 16kHz sampling rate
- Audio output: 24kHz
- Connection method: Stateful long connection based on WebSocket
- Supports 70+ languages
Because there are no intermediate conversion steps, latency is significantly reduced, and it can capture information like tone and speaking speed that text cannot convey.
What It Can Do
Barge-in: You can interrupt the AI while it is speaking, re-ask a question, or change direction — traditional TTS cannot do this; you have to wait until a sentence finishes before speaking the next one.
Tool calling: It can call functions and Google Search, meaning the model can look up information and execute actions in real time during a voice conversation.
Dual transcription: The content of both sides of the conversation can be transcribed simultaneously, making it easy to record and analyze.
Two deployment modes: Server-to-server (suitable for backend systems like customer service bots) and client-to-server (suitable for user-facing voice applications).
Benchmark Results
On ComplexFuncBench Audio, a benchmark specifically designed to test voice models in complex function-calling scenarios, Gemini 3.1 Flash Live scored 90.8%.
Real-World Impact
One of the biggest obstacles for voice AI is that pause-like feeling, similar to waiting for a phone call to connect. Reducing latency and supporting barge-in are essential conditions for making voice interaction truly close to natural conversation.
This model is available to developers through the Gemini Live API in Google AI Studio and has also been integrated into Gemini Live (available to users in over 200 countries).
Google's strategy on the voice front, from Gemini 2.5 Flash Native Audio to the current 3.1 Flash Live, is very consistent: treat voice as a first-class citizen, rather than just wrapping a TTS layer around an LLM.
Sources: Gemini 3.1 Flash Live: Making audio AI more natural and reliable (Google Blog); CocoLoop, Google Releases Gemini 3.1 Flash Live: A Real-Time Multimodal Voice Model (MarkTechPost); Gemini Live gets its biggest upgrade yet with Gemini 3.1 Flash Live (9to5Google)