Google's Gemini 3.8 Brings Programmable Voices and Sends TPUs Into Orbit

Google’s latest weekly AI release spans production tools and long-range research: two text-to-speech models, a video avatar for enterprise agents, new study features in Gemini Notebook, and an orbital test of TPU hardware. The near-term releases add programmable voice, visual agents, and source-grounded tutoring, while Project Suncatcher tests whether machine-learning infrastructure can operate in space.

Prompted voices, line by line

Google’s new TTS models, Gemini 3.8 Flash TTS and Flash-Lite TTS, generate voices from natural-language descriptions rather than limiting developers to presets. Both are available through Google AI Studio, the Gemini API, Gemini Enterprise, Gemini Notebook, and Google Vids.

Flash targets workloads that require detailed creative direction, while Flash-Lite prioritizes high-volume, cost-sensitive uses such as dubbing and voice agents. Developers can prompt for a character, accent, speaking style, or regional dialect across more than 100 languages and dialects. Google also provides a library of more than 2,000 voices.

Inline stage directions control pacing, emotion, accent shifts, and vocal effects at the line level. Scripts can include cues such as <laughs>, <sigh>, and |mhm|, or assign distinct voices to a two-person conversation while preserving turn-taking and speaker consistency across long recordings.

Controls exposed to builders

• Generate a character voice from a text prompt without reference audio.
• Replicate a voice from about 30 seconds of audio after consent verification.
• Adjust pacing, emotion, accents, and vocal cues within a script.
• Produce two-speaker dialogue from one prompt.
• Maintain a speaker’s voice across audiobooks, podcasts, and other long-form audio.
• Attach SynthID watermarks and C2PA provenance credentials to generated output.

Google reports a score of 71.4 and first place on Hume AI’s Voice Design Benchmark, along with the top two positions on its Overall Quality Index for Flash and Flash-Lite. Those vendor-reported rankings provide a quality signal, although production evaluations will still need to account for latency, cost, language coverage, and consistency.

Voice replication in AI Studio is unavailable in Illinois, Texas, the European Economic Area, the United Kingdom, Switzerland, and India. Teams planning to use replication should confirm regional access before making it a dependency.

An avatar that works while it talks

Gemini 3.8 Live with Live Avatar combines streaming speech and video generation to create enterprise agents with synchronized speech, facial expressions, and conversational turn-taking. Google first previewed the feature at Cloud Next 2026 and now offers it generally through Gemini Enterprise.

Gemini 3.8 Live accepts audio, images, video, and text within a shared context window of up to 128,000 tokens, which determines how much input the model can retain during a session. Live Avatar returns audio, video, and text with an output limit of 24,000 tokens. It supports lip synchronization and turn-taking in 97 languages while adjusting to changes in speech without breaking video alignment.

Asynchronous tool calling allows an avatar to query systems or update records while continuing the conversation. A claims-intake agent, for example, could populate a form in the background while a customer describes damage on camera, rather than pausing the session for each external request.

Google limits customers to a curated library of prebuilt avatars by default. Creating a custom avatar requires enterprise allowlisting and verification, controls intended to reduce impersonation and deepfake abuse.

Notebook talks through your sources

Gemini Notebook, formerly NotebookLM, is adding Interactive Learning Overviews that combine source-based summaries, infographics, quizzes, and flashcards. The feature appears under Reports and is scheduled to reach users over the next few weeks.

The update also adds short-answer, multiple-select, and fill-in-the-blank questions, plus shareable 60-second Video Overviews in more than 80 languages. These formats give students and researchers several ways to test recall without moving material into a separate study application.

Live Chat on Android and iOS adds real-time spoken conversations with a notebook in nearly 100 languages. Users can ask follow-up questions, request step-by-step explanations, and interrupt an answer while the notebook is speaking. Access begins with Google AI Ultra subscribers, followed by Google AI Pro subscribers and other users.

Because Notebook grounds responses in sources selected by the user, Live Chat can serve as a hands-free interface for a defined research corpus. Its usefulness will depend on source quality and the model’s citations, especially when a notebook contains conflicting or incomplete documents.

Four TPUs take the orbital test

Project Suncatcher is Google’s research program for evaluating whether AI hardware could eventually operate in space. Its first orbital mission will test four Tensor Processing Units, or TPUs, aboard a prototype satellite rather than attempt a full orbital data center.

The refrigerator-sized craft, called MVP, draws about 1 kilowatt of solar power and runs workloads that include short bursts of Gemini queries. Each burst lasts roughly 15 minutes before the hardware pauses to cool. Google expects to operate the satellite for about a year, although the craft could remain in orbit for as long as six years.

A SpaceX Falcon 9 will launch MVP from Vandenberg Space Force Base as part of the Transporter-18 rideshare mission. Google built the satellite with Planet, an Earth-imaging company experienced in deploying compact spacecraft.

What the flight must prove

• Can the TPUs survive launch vibration and sustained acceleration of up to 10g?
• Can individual components tolerate transient forces between 50g and 100g?
• Can heat pipes and radiators remove TPU heat through passive cooling in a vacuum?
• Will the chips continue operating after months of exposure to low-Earth-orbit radiation?

Google’s energy case for orbital compute rests on stronger, longer access to sunlight. The company estimates that suitable satellites could generate up to eight times as much solar power as comparable systems on Earth, where weather, the atmosphere, and the day-night cycle reduce output.

Google’s preprint sketches a larger system built from TPU-equipped satellites, solar arrays, and free-space optical links, which transmit data between spacecraft with lasers. One proposed configuration uses 81 satellites within a cluster roughly 1 kilometer in radius. A separate two-satellite experiment planned for 2027 will test whether those laser links can deliver the bandwidth needed to distribute machine-learning workloads.

Suncatcher’s first flight remains a hardware and thermal-management experiment with substantial engineering and operational hurdles ahead. Across the broader release, Google is concentrating its production work on programmable speech, visual agents, and source-grounded conversation while testing a possible infrastructure path much further out.