Speak naturally
Press a hotkey and say the message, note, draft, or instruction you would otherwise type.
Gemma 4 voice dictation
Voice dictation that feels closer to typing.
Cue is a desktop voice AI for Mac and Windows. A May 2026 Gemmaverse case study documents a Cue dictation benchmark in which cloud speech-to-text was followed by a local Gemma 4 E4B text-polish step.
Featured in Google DeepMind's Gemmaverse. The case study reports results from a 227-sample Apple Silicon benchmark, as of May 2026.
Median polish latency876 → 488 ms44% lower median polish latency in the published benchmark.
Feature usage+30%Reported after the faster local pipeline was introduced.
Why the dictation pipeline matters
Most people do not search for a model; they search because speaking into a text field still feels slower or rougher than typing. A useful desktop dictation loop has to recognize the speech, clean it up, and place it back into the task before the user loses their train of thought.
Polish the text
In the published benchmark path, cloud speech-to-text produced raw text and Gemma 4 E4B polished that text locally on Apple Silicon.
Continue the task
The result lands in the current app, where it can become the reply, document, prompt, or follow-up.
What the published case study measures
Google DeepMind's Gemmaverse profile describes replacing Cue's cloud text-polish step with a local Gemma 4 E4B polish step while cloud speech-to-text remained in the path. It reports 44% lower median polish latency, a 30% increase in feature usage, and zero marginal inference cost for that local polish step. Those published May 2026 results describe the measured path, not a promise that every device, release, or task will see the same number.
The practical point is simple: a voice tool earns a place in a daily workflow when it responds quickly enough to use for the small things—replying in Slack, drafting an email, taking a note, or getting rough text into a document. That is why Cue treats dictation as an entry point to work across your apps, rather than a separate recording destination.
For a broader evaluation plan, use the public voice agent benchmark guide to separate task completion, spoken tool use, browser action, dictation, latency, and meeting understanding. The open landscape is a separate research resource; it does not reproduce Cue's private 227-sample production benchmark.
Start with the job you need to finish
Use the same desktop workflow for voice typing in a text field, a longer live transcript, or a meeting-note follow-up. Each route starts with spoken input and ends with a useful next action.
Try Cue's desktop dictation.
Cue is free to start on Mac and Windows. Press a hotkey, speak, and keep working in the app already in front of you.