Google debuts Gemini 3.5 Transcribe, a smarter speech-to-text model
Two Google model releases dominate the AI news cycle this afternoon — neither a weight-dump spectacle, both quietly aimed at making AI's messier inputs tractable. One turns raw audio into clean, structured text; the other turns a stream of glucose readings into transferable medical insight. Here's the rundown.
Google opens Gemini 3.5 Transcribe to developers, a speech-to-text model built to clean up messy audio. Announced Wednesday in public preview for the Gemini API and the Gemini Enterprise Agent Platform, Gemini 3.5 Transcribe is Google's most precise speech-to-text model yet, engineered for real-time and pre-recorded transcription that survives background noise, rapid-fire jargon, and the verbal detritus of actual human speech. It powers the Rambler feature on Gboard for Android, will reach Chrome as a "talk to type" capability, and already backs voice dictation in the Gemini app on macOS.
The model is the successor to Chirp 3, and it shows in the numbers: Google cites a 4.0% word error rate for streaming and 2.6% for non-streaming transcription as measured by Artificial Analysis, with time-to-final-transcription improved by 70%. What makes it interesting beyond the raw accuracy is how it handles the stuff that normally wrecks speech models — self-corrections ("let's meet Tuesday—no, Wednesday"), filler words, and industry vocabulary through a customizable list — while attributing speech to up to three speakers and detecting over 85 languages. It also routes function calls to other Gemini models, so dictation can hand off tasks like image generation and file analysis. Developers get two pathways: a Live API for sub-second streaming and an Interactions API for recorded meetings and call logs with word-level timestamps and speaker attribution. For teams building voice agents, real-time captioning, or post-call analytics, this is a serious, ready-to-integrate upgrade rather than a research demo.
Google's GlucoFM brings a foundation model to continuous glucose monitoring. Google Research introduced GlucoFM, a lightweight, self-supervised foundation model for continuous glucose monitoring that separates slow glycemic trends from short-term deviations into two streams before combining them. Pre-trained on roughly 109,000 hours of unlabeled CGM data across 477 participant records, the model's frozen representations beat prior baselines across four cohorts and seven metabolic tasks — diabetes risk, insulin resistance, beta-cell dysfunction, hypoglycemia, and more — lifting average PR-AUC by about 4 points versus the strongest CGM-specific baseline retrained on the same data. On postprandial glucose forecasting it posted the lowest mean absolute error of the field, and it transfers well between clinical cohorts and adapts with strikingly few labels.
The significance is less the headline accuracy than the economics of medical AI it points toward. High-quality clinical labels are scarce and costly, yet GlucoFM claims strong few-shot performance — meaningful results with as little as one labeled participant per class — which could let downstream tools be built on far less expensive annotation. It's a natural extension of Google's broader push into wearable health intelligence we've tracked before — we covered the company's insulin-resistance trend feature on the Pixel Watch 5 and Fitbit in early August in Pixel Watch 5 tracks insulin resistance — a wearable first. The next step for GlucoFM is larger and more diverse pre-training data, and modeling that looks beyond single 24-hour windows toward native multi-day dynamics.
Both launches reframe a familiar debate — transcription sounds solved, glucose tracking sounds niche. Where do you think speech-to-text models like this actually break down in daily use? Tell us in the comments.
