Google launches Gemini 3.8 Live for real-time voice

Google has added Gemini 3.8 Live, Gemini 3.8 Live Extended Thinking and Gemini 3.5 Transcribe to its tools for building real-time voice applications. The goal is to create agents that can hold conversations, understand what is happening and execute tasks without interrupting the dialogue.

Until now, many voice applications have worked as a chain of steps: one model converted audio into text, another generated a response and an additional system turned it back into speech. The new Live models aim to reduce that complexity through direct voice interaction.

Agents that speak and act at the same time

Gemini 3.8 Live can maintain a conversation while performing actions in the background. For example, a customer service agent could check an order's status through an API while continuing to respond to the user, instead of going silent during the search.

Its main features include:

• Asynchronous function calls, to run tools and APIs while the spoken response continues.
• Visual context, with the ability to combine what the user says with live images or video.
• Greater accuracy when reading alphanumeric data, such as confirmation codes, claim numbers or technical references.
• Support for more than 97 languages, with consistent accents.
• Incremental updates, which make it possible to combine real-time audio with structured data as it arrives.

The Gemini 3.8 Live Extended Thinking version adds a configurable reasoning mode. It can spend more time solving complex, multi-step requests while responding or explaining its progress within the conversation. Google says it ranks first in Artificial Analysis's ranking for this type of task.

The practical difference matters: a voice agent would not have to limit itself to answering simple questions. It could also arrange a return, review several insurance conditions or guide you step by step through a technical process while keeping the conversation open.

More accurate and adaptable transcription

Google also highlights Gemini 3.5 Transcribe, a model specialized in converting speech into text. It supports more than 85 languages and recorded an average word error rate of 4.0% on streamed audio and 2.6% on non-streamed audio. In this metric, a lower percentage means fewer transcription errors.

The model includes features designed for real-world environments:

• Automatic language switching, even within the same sentence.
• Custom vocabulary, with up to 1,000 terms for company names, professional jargon or uncommon words.
• Intelligent transcription mode, which organizes the text, corrects speech errors and removes filler words.

This could be useful for subtitles with less than one second of latency, call analysis, contact center assistants and tools that turn meetings or conversations into usable text.

Through the Interactions API, Gemini 3.5 Transcribe can also process audio files of up to one hour, with structured timestamps and speaker separation. This means the model indicates not only what was said, but also when and by whom.

What changes for application developers

The models are available through the Gemini API and Google AI Studio. Google also provides examples on GitHub and access to other parts of its audio platform, including speech translation in more than 70 languages, voice generation with Gemini 3.1 Flash TTS and music creation with Lyria 3.5.

For you, the most visible change will come when a voice application stops feeling like a phone menu. The response can continue while the system retrieves data, interpret names and codes more accurately, and combine voice, text and images. What needs watching now is accuracy in long conversations and noisy situations, because that is where it will become clear whether these agents are genuinely useful or merely sound fluid in demonstrations.