Google quietly shipped two significant Gemini model updates within 48 hours this week, and both are aimed squarely at developers building production applications. Gemini 3.5 Transcribe, introduced Aug. 26, is the company's most precise speech-to-text model to date, according to Google's official blog. Gemini Omni 1.1 Flash, announced Aug. 27, brings professional-grade controls to generative video, including scene extension up to 40 seconds and 4K upscaling. Both models are available through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform.
Gemini 3.5 Transcribe: speech-to-text that cleans up after itself For more context on this story, see our ongoing AI news.
What distinguishes Gemini 3.5 Transcribe from conventional speech recognition, Google says, is that it converts raw audio directly into polished, formatted text rather than a raw transcript. The model handles background noise, complex jargon, and disfluency cleanup natively — it removes filler words like "ums" and "ahs," resolves self-corrections such as "let's meet Tuesday—no, Wednesday," and auto-formats the output.
The benchmark numbers Google cited are notable. As measured by Artificial Analysis, the model achieves an average word error rate (WER) of 4.0 percent for streaming use cases and 2.6 percent for non-streaming audio. On the FLEURS multilingual benchmark across top languages, it posts a 5.50 percent WER in streaming mode and 5.04 percent non-streaming, improving on Google's previous transcription model, Chirp 3. Time to final transcription improves by 70 percent compared with Chirp 3, again as measured by Artificial Analysis.
The model ships in two variants for two workflows. For live applications, `gemini-3.5-transcribe-live` delivers continuous bidirectional streaming with sub-second latency through the Live API, targeting voice agents and real-time captioning. For recorded audio — meetings, call logs, archives — `gemini-3.5-transcribe` offers speaker attribution for up to three speakers (with support for more in experimental mode) and word-level timestamps via the Interactions API.
Other capabilities include automatic detection and transcription across more than 85 languages with accent and dialect handling, and custom vocabulary support so the model adapts to specialized jargon and unique spellings. Google also gave the model eyes of a sort: through function calling, it can delegate tasks like image generation or file analysis to other Gemini models — a feature currently available in the Gemini app on macOS.
Gemini Omni 1.1 Flash: video generation grows a timeline
The second launch addresses the most common complaint about AI video: control. Gemini Omni 1.1 Flash is a production-focused update that makes generative video steerable in ways its predecessors were not, with Google DeepMind product managers Anish Nangia and Alisa Fortin describing the release as making Omni 1.1 "production-ready for professional use."
Scene extension is the headline feature. Developers can now continue generating footage seamlessly from an existing clip, with the model analyzing up to 10 seconds of prior context — a leap from previous models that referenced only the final second. Videos can be extended in 10-second increments up to a cumulative length of 40 seconds, improving visual consistency and narrative adherence for longer stories.
The update also adds first-and-last-frame interpolation, letting developers specify start and end frames to create smooth, deliberate camera movements and transitions between shots. For delivery, finished projects can be upscaled to 4K resolution, while a 360p preview mode lets creators iterate on prompts quickly and cheaply before committing to final renders.
From API to everyday surfaces
Google is also wiring both models into its consumer products, which is where its distribution advantage shows. On Android, the new "Rambler" feature in Gboard uses 3.5 Transcribe to turn spoken thoughts into well-formatted text while filtering filler words — users can even make edits by voice. The model also powers dictation in the Gemini app on macOS, works alongside screen context in Google Antigravity, and is being integrated into Chrome.
For developers, the two launches read as one strategy: Google is pushing Gemini from a chatbot you talk to into infrastructure that sees, hears, and produces media. With OpenAI, ElevenLabs, and a cohort of video startups competing in the same territory, the 2.6 percent word error rate and 40-second coherent scenes are Google's opening bid for the production workloads that actually generate revenue — voice agents, captioning pipelines, and creative tools. Developers can start building with both models in Google AI Studio today.
Why word error rate still matters
Word error rate sounds like an academic statistic, but in production it is the difference between a voice product that works and one that frustrates. Every misunderstood utterance in a voice agent breaks a task: a misheard order ID triggers a failed lookup, a garbled address sends a package to the wrong city. That is why the gap between a 2.6 percent WER on non-streaming audio and the high-single-digit rates that were common a generation of models ago compounds into an order-of-magnitude difference in usability.
The distinction between the two modes Google reported also matters for architects. Streaming transcription — the 4.0 percent WER figure — trades some accuracy for sub-second latency, which interactive use cases like live voice agents require. Batch transcription of recorded audio runs slower but reaches 2.6 percent, which is why post-call analytics and archive processing can afford to be more demanding. Google's decision to expose both as separate model endpoints rather than one adjustable setting acknowledges that developers build different products on either side of that tradeoff.
A tiered workflow for video production
The Omni 1.1 Flash update is equally pragmatic about how creative work actually happens. Generative video has been expensive to iterate with: every prompt experiment meant a full render. The new 360p preview mode separates exploration from delivery, letting creators cycle through prompt variations cheaply and only paying for 4K upscaling on the shots that survive review.
The scene extension mechanics address the coherence problem that has kept AI video in the clip-sized ghetto. By analyzing 10 seconds of prior context instead of only the final frame, the model can extend a scene in 10-second increments up to 40 seconds while keeping characters, lighting, and visual style consistent. Combined with first-and-last-frame interpolation — which gives editors deterministic control points, the way in and out markers work in conventional editing — AI-generated footage starts to fit into real post-production pipelines rather than replacing them wholesale.
The competitive picture
Both launches position Google as infrastructure rather than destination. In speech, Google is taking on specialized transcription vendors and the voice stacks of its cloud rivals by bundling state-of-the-art accuracy into the Gemini API its developers already use. In video, it is competing in a market crowded with well-funded startups by emphasizing control and workflow integration over raw spectacle.
The distribution advantage may be the deciding factor. The same transcription model that developers can call from the Gemini API already powers Gboard's Rambler dictation on Android, the Gemini app on macOS, Google Antigravity, and Chrome — meaning Google can amortize model development across billions of user interactions while collecting the diverse real-world audio that keeps improving it. For developers choosing a speech or video stack in 2026, that flywheel is now part of the pitch.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
