Summary
Google has released Gemini 3.8 Live, Gemini 3.8 Live Extended Thinking and Gemini 3.5 Transcribe for developers building real-time voice applications. The models add speech-to-speech task execution, visual context and transcription across more than 85 languages.
Google has released three audio models for developers building real-time voice applications: Gemini 3.8 Live, Gemini 3.8 Live Extended Thinking and Gemini 3.5 Transcribe. The models are available through the Gemini API and Google AI Studio, with the Live models offered through Google’s Live API.
The release combines native speech-to-speech interaction with a dedicated speech-to-text system. Google says Gemini 3.8 Live can reason and execute tasks while maintaining a conversation, while Gemini 3.5 Transcribe is designed for low-latency transcription across more than 85 languages.
Voice agents that can act during a conversation
Gemini 3.8 Live is intended for applications in which an agent needs to listen, respond and perform actions without breaking the flow of dialogue. Google lists asynchronous function calling as a core capability: an application can execute API or tool calls in the background while the model continues streaming an audio response.
The models can also use live visual inputs as context, allowing an agent to combine what a person says with what is visible to the system. Other listed capabilities include parsing alphanumeric information such as confirmation codes, claim numbers and technical data, and supporting more than 97 languages with accent consistency.
Gemini 3.8 Live Extended Thinking adds configurable reasoning for complex, multi-step requests. It can work through that reasoning in the background while responding to the user or narrating progress in the ongoing conversation.
In a conventional cascaded voice system, separate speech-recognition, language-model and speech-synthesis stages pass information from one component to the next. Google presents the new Live models as a more streamlined alternative to that approach, because they are built for native speech-to-speech interaction while also supporting tools, structured information and visual context.
Google says the Live models are priced at $0.005 per minute for audio input and $0.018 per minute for audio output. They can also be accessed through integration partners including Agora, Fishjam, LiveKit, LangChain, Pipecat, Vercel and Vision Agents, which provide media-streaming infrastructure for deployment.
Transcription for multilingual and specialised audio
Gemini 3.5 Transcribe converts streamed speech into text. Google reports an average Word Error Rate of 4.0% for streaming transcription and 2.6% for non-streaming transcription. Word Error Rate measures the proportion of transcription errors relative to the words in the reference transcript; lower values indicate fewer errors.
The model supports automatic code-switching, meaning it can handle speakers moving between languages within or across sentences without manual configuration. Developers can also supply a custom vocabulary list of up to 1,000 terms to improve recognition of domain-specific language, uncommon jargon, company names and proper nouns.
A smart transcription mode can produce more polished text by adding structured formatting, removing filler words and handling self-corrections. Google lists potential uses including real-time audio analytics, call-centre agents and sub-second captioning.
Through the Interactions API, Gemini 3.5 Transcribe can also process audio files of up to one hour and provide structured timestamps and speaker labelling. This extends the model beyond live conversations to applications that need searchable or speaker-separated transcripts.
A broader audio development stack
The new models sit alongside Google’s other audio tools in the Gemini API. These include 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.
For developers, the practical change is access to separate building blocks for conversational speech, transcription, translation and generated audio. The Live models target interactive agents that can respond and call tools in real time, while Transcribe provides a specialised path for turning live or recorded speech into structured text.