Creating a voice application that listens, reasons, and takes action while maintaining a natural conversation usually requires connecting several models and services. Google is looking to simplify that process with new audio models available through the Gemini API and Google AI Studio.
The company introduced Gemini 3.8 Live, Gemini 3.8 Live Extended Thinking, and Gemini 3.5 Transcribe—a combination designed for conversational agents, real-time transcription, and more complete voice experiences.
Gemini 3.8 Live brings reasoning to voice conversations
The Gemini 3.8 Live models let you create agents that do more than respond with audio. They can also reason, access tools, and carry out tasks without breaking the flow of the conversation.
The difference matters. In a traditional architecture, an application may need one model to transcribe audio, another to generate the response, and a third to convert text into speech. Gemini is aiming for a native voice-to-voice experience that reduces that complexity.
Its main features include:
- Asynchronous function calls: the agent can run APIs or tools in the background while continuing to deliver an audio response.
- Visual context: it can combine what the user says with live visual inputs to better understand the situation.
- Alphanumeric accuracy: it is designed to interpret confirmation codes, claim numbers, and technical data more accurately.
- Multilingual support: it covers more than 97 languages and aims to preserve accent consistency.
- Incremental updates: it can combine real-time audio with structured data to deliver more relevant responses.
Imagine a support agent that talks with you, checks an order’s status through an API, and responds by voice while waiting for the data. It would not have to fall silent during every external call. That continuity can make the interaction feel much more natural.
Configurable reasoning for complex requests
Gemini 3.8 Live Extended Thinking adds a configurable reasoning mode. The model can work on multistep tasks in the background while responding or explaining its progress within the main conversation.
Google says this version ranks first on the Artificial Analysis reasoning leaderboard among the models evaluated. As always, this kind of position depends on the date, methodology, and tests used, so it is worth reviewing the benchmark before making technical decisions.
Gemini 3.5 Transcribe turns voice into text accurately
For applications that need real-time text, Google also introduced Gemini 3.5 Transcribe. It is a specialized speech-to-text model that supports more than 85 languages.
According to Google, it achieved an average Word Error Rate (WER) of 4.0% in streaming transcription and 2.6% in non-streaming transcription. WER measures the proportion of incorrectly transcribed words, although the actual result can vary depending on background noise, the microphone, the accent, and the type of conversation.
The model includes features that are especially useful for voice products:
- Automatic language switching: it recognizes language changes within the same sentence or between sentences without manual configuration.
- Custom vocabulary: it lets you prioritize specialized terms, company names, proper names, and industry jargon through a
custom_vocabularylist of up to 1,000 terms. - Intelligent transcription mode: it produces cleaner text with structured formatting, autocorrections, and filler-word removal.
These capabilities can be useful for subtitles with under one second of latency, call center agents, conversation analysis, and accessibility tools. You can also use the model through the Interactions API to transcribe audio files of up to one hour, with structured timestamps and speaker separation.
A simpler audio architecture for developers
Google’s proposal is not limited to a single model. The company is bringing together several components for building audio products through the Gemini API:
- Gemini 3.5 Live Translate: voice-to-voice translation in more than 70 languages.
- Gemini 3.1 Flash TTS: configurable voice generation.
- Lyria 3.5: music generation designed for production environments.
To get started, developers can try the models in Google AI Studio Live, clone sample applications from GitHub, or use the Live API agent capabilities guide.
From a technical perspective, the main appeal is reducing the need for cascading architectures. Instead of manually coordinating speech recognition, reasoning, tool calls, and audio synthesis, an application can delegate more steps to a native voice model. That does not eliminate the need to design latency, permissions, observability, and error handling carefully, but it can speed up development.
Voice is no longer just an interface for sending commands. With these models, it can become a complete layer for holding conversations, retrieving information, interpreting images, and taking action.
The challenge will be seeing how these promises hold up outside demos: with background noise, interruptions, accents, long conversations, and sensitive data. Even so, the direction is clear. Voice applications no longer have to be limited to answering questions; they can become agents capable of supporting a task from beginning to end.