Google DeepMind Launches Gemini 3.8 Live and 3.5 Transcribe for Voice AI
New speech-to-speech and transcription models enable more intelligent conversational experiences
1Breakthrough in conversational AI
Google DeepMind released Gemini 3.8 Live and 3.8 Live Extended Thinking models, marking a significant advancement in native speech-to-speech technology. These models can perform tasks while maintaining dialogue flow, with Extended Thinking handling complex multi-step reasoning. The models ranked #1 on Artificial Analysis' Speech-to-Speech leaderboard, demonstrating superior performance in real-time conversational AI applications.
2Enhanced transcription capabilities
The Gemini 3.5 Transcribe model offers highly precise speech-to-text conversion across 85+ languages. Released last month, it achieves an impressive Word Error Rate of 4.0% for streaming and 2.6% for non-streaming transcription. This specialized model serves as a robust listening engine for various voice applications, including sub-second captioning and call center operations.
3Technical specifications
Gemini 3.8 Live models are available through the Live API, priced competitively at $0.005 per minute for audio input and $0.018 per minute for output. The models provide a streamlined alternative to cascaded architectures, supporting configurable thinking for complex reasoning tasks. Developers can integrate these models with various media streaming partners including Agora, Fishjam, and LiveKit.
4Developer accessibility
Google has made these models accessible through multiple channels including the Gemini API, Google AI Studio, and GitHub for example apps. The Interactions API allows transcription of audio files up to one hour with structured timestamps and speaker labeling. Developers can quickly start experimenting with these tools through ai.studio/live and various integration partners.
5Real-world applications
The new models enable creation of voice agents capable of reasoning and task execution while maintaining conversation flow. Potential applications include real-time audio analytics, call center automation, and interactive voice interfaces. The technology supports stateless tasks like instant captioning while offering extended capabilities for complex dialogue management.
6Performance benchmarks
Gemini 3.5 Transcribe sets new standards with its 4.0% WER in streaming mode, significantly improving upon previous transcription technologies. The 3.8 Live models demonstrate superior performance in maintaining dialogue continuity, particularly in Extended Thinking mode which handles background reasoning without disrupting primary conversation flow.
7Integration ecosystem
The release comes with extensive support through integration partners like LangChain, Pipecat, and Vercel that handle media streaming infrastructure. This ecosystem approach allows developers to deploy voice applications at scale while leveraging industry-leading performance. The models complement Google's existing speech and music generation offerings in the Gemini API.
8Future of voice interfaces
These advancements represent a significant step toward more natural and intelligent voice interactions. By combining high-accuracy transcription with advanced speech-to-speech capabilities, Google DeepMind is enabling developers to create next-generation conversational AI applications that blur the line between human and machine interaction.