Google is expanding its Gemini lineup for real-time voice applications with two new audio models designed to make AI agents more capable during live conversations.

The company has introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking for developers building voice-first applications, alongside Gemini 3.5 Transcribe, a dedicated speech-to-text model designed for fast and accurate transcription. The models are available through the Gemini API and Google AI Studio.

The announcement marks another step in Google’s push toward AI agents that can do more than simply listen and respond. With the new Live models, developers can build systems that reason, call tools, process visual information and execute tasks while continuing a natural voice conversation.

Gemini 3.8 Live is built for real-time AI agents

Gemini 3.8 Live is Google’s latest native speech-to-speech model, allowing developers to create conversational agents that can perform tasks without breaking the flow of a spoken interaction.

One of the key additions is asynchronous function calling. Instead of stopping the conversation while an API or tool call is being processed, an agent can execute those operations in the background while continuing to stream an audio response to the user.

That could make a meaningful difference for applications where an AI assistant needs to interact with external systems. For example, an agent could continue talking with a customer while it retrieves information or performs another operation through a connected tool.

The models can also use visual context, allowing voice agents to ground their responses in live visual input. This opens the door to experiences where an AI can understand both what a person is saying and what they are showing it.

Google is also highlighting improved handling of alphanumeric information, including confirmation codes, claim numbers and other technical data where accurately distinguishing individual characters can be important.

Multilingual support covers more than 97 languages, with Google also emphasizing consistency in accents. The models can additionally provide incremental content updates, allowing real-time audio responses to be combined with structured information as an interaction progresses.

For more complicated tasks, Gemini 3.8 Live Extended Thinking adds configurable reasoning. The model can work through multi-step problems in the background while continuing the conversation and, where appropriate, responding or narrating its progress.

Google says Gemini 3.8 Live Extended Thinking ranked first on Artificial Analysis’ Speech-to-Speech leaderboard, although such benchmark results are based on the methodology and evaluation set used by the organization.

Google is also targeting the cost of real-time voice AI

Both Gemini 3.8 Live models are available through Google’s Live API.

Google lists pricing at $0.005 per minute for audio input and $0.018 per minute for audio output. The company says the pricing is intended to make it possible for developers to scale real-time voice applications while maintaining the performance needed for conversational agents.

Google is also working with several infrastructure and developer platforms to make the Live API easier to deploy in production. Partners include Agora, Fishjam, LiveKit, LangChain, Pipecat, Vercel and Vision Agents.

That ecosystem matters because building a production voice agent involves considerably more than the underlying AI model. Real-time audio streaming, media infrastructure, tool integrations and application logic all need to work together without introducing noticeable delays.

Gemini 3.5 Transcribe focuses on accurate speech-to-text

Alongside the Live models, Google is highlighting Gemini 3.5 Transcribe, its dedicated speech-to-text model for low-latency transcription.

Google says the model supports more than 85 languages and achieved an average Word Error Rate of 4.0% for streaming transcription and 2.6% for non-streaming transcription in its reported evaluation.

The model includes automatic code-switching, allowing it to handle users switching between languages within a conversation without requiring developers to manually configure each language change.

Developers can also provide a custom vocabulary of up to 1,000 terms. That can help the system recognize specialized terminology, uncommon jargon, company names and proper nouns that might otherwise be difficult for a general speech-recognition system to identify correctly.

Another feature is Smart Transcription mode. Rather than simply converting every spoken word into text, it can produce cleaner, reader-ready transcripts with structured formatting, self-corrections and filler-word removal.

That makes Gemini 3.5 Transcribe useful beyond conversational agents. Google specifically points to applications including sub-second captioning, call-center systems and real-time audio analytics.

The model is also available through the Interactions API for transcribing audio files of up to one hour, with structured timestamps and speaker labeling.

Google’s audio toolkit is becoming broader

Gemini 3.8 Live and Gemini 3.5 Transcribe join an increasingly broad collection of audio capabilities available through the Gemini API.

Google already offers Gemini 3.5 Live Translate for speech-to-speech translation across more than 70 languages, Gemini 3.1 Flash TTS for configurable speech generation and Lyria 3.5 for music generation.

Together, these models give developers building blocks for different parts of an audio experience: understanding speech, generating responses, translating conversations and creating audio or music.

For developers, the bigger shift is that voice AI is increasingly being treated as an agent interface rather than simply a speech-recognition feature. Gemini 3.8 Live is designed to let an agent listen, reason, use tools and respond while maintaining a continuous conversation, while Gemini 3.5 Transcribe provides a dedicated transcription layer for applications that need fast and precise speech recognition.

Developers can try the new Live models through Google AI Studio or build applications using the Gemini API and Live API.