Google DeepMind has opened access to two new audio models for developers: Gemini 3.8 Live, a speech-to-speech system, and its heavier sibling, 3.8 Live Extended Thinking, both available now through the Gemini API and Google AI Studio.
The headline capability is that these models can act while they talk. Google’s product team, in a developer blog post credited to product manager Alisa Fortin and technical staff member Thor Schaeff, describes asynchronous function calling: the model can fire off an API request or tool call in the background and keep streaming audio to the user rather than going silent while it waits on a result. That is a structural change from cascaded voice pipelines, which typically transcribe speech to text, run a separate reasoning step, then synthesize a reply, introducing pauses at each handoff.
Google also claims 3.8 Live Extended Thinking now ranks first on Artificial Analysis’s Speech-to-Speech leaderboard, a third-party benchmark rather than an internal one. The post does not say by what margin or against which competing models, so the claim is verifiable only by checking Artificial Analysis directly rather than taking Google’s summary at face value.
Beyond the reasoning upgrade, the 3.8 Live family adds visual context grounding (letting an agent reference what a user is showing it, not just saying), alphanumeric precision for reading back confirmation codes and account numbers, and coverage across more than 97 languages. Google prices the models at $0.005 per minute of audio input and $0.018 per minute of audio output through the Live API, and says seven infrastructure partners, including LiveKit, LangChain, Pipecat and Vercel, already support the models for teams that do not want to build their own media-streaming layer.
The second release, Gemini 3.5 Transcribe, is not new: Google says it shipped last month and is now being packaged alongside the Live models as part of a broader audio suite. Google reports an average word error rate of 4.0 percent in streaming mode and 2.6 percent non-streaming across more than 85 languages, and says it switches between languages mid-sentence on its own. Developers can also supply up to 1,000 custom terms to steer it toward proper nouns and jargon. A separate Interactions API route handles longer files, up to one hour, with timestamps and speaker labels attached.
None of these figures come with a competitor comparison. Google does not publish, in this post, how its pricing or error rates stack up against rival real-time voice APIs, so the “industry-leading performance” framing in its own announcement should be read as a vendor claim pending independent testing rather than an established fact.
What matters for builders is the shift in default architecture. A team assembling a voice agent today has to choose between a cascaded stack, transcription, a language model, then text-to-speech stitched together, and a native speech-to-speech model that handles turns and tool calls in one pass. Google is betting the async function-calling model removes enough latency that cascaded pipelines become the legacy option for anything beyond simple dictation.
Teams already running voice agents on cascaded architectures should benchmark call-handling latency and confirmation-code accuracy against 3.8 Live before their next infrastructure renewal, since a wrong digit in a claim number or account lookup is the kind of error that erodes trust in a voice interface fastest.
Reported by Google’s developer blog (blog.google), post published September 15, 2026.
