Gemini 3.5 Transcribe can tidy your spoken prompts. Check what it changes.
Maestro Brief · Published by Maestro Mojo
2026-08-27
Maestro’s take
Talking through a bug is easy. Turning that ramble into a clear task is the useful part.
Google’s new Gemini 3.5 Transcribe makes that a concrete developer feature. But cleaner wording is not automatically a faithful record. Treat it as a draft you check, especially before handing it to an agent that can change code.
A polished mistake is still a mistake.
Published August 27, 2026 · Maestro’s opinion on model news.
TL;DR
Google introduced Gemini 3.5 Transcribe on August 26. Developers can try it in public preview for live speech or recorded audio.
The announcement also describes voice input in AI Studio’s app-building mode and screen-aware transcription in Antigravity, with permission. General “talk to type” support in Chrome is coming soon, not a feature to assume is available today.
This is speech-to-text. Building or changing your app still requires a coding agent or another workflow.
The setting that matters
Google’s live API documentation offers two modes:
| Mode | What it is designed to do |
|---|---|
VERBATIM—the default |
Preserve spoken words, including fillers and false starts. |
SMART |
Remove fillers, resolve spoken corrections, and format the text. |
Neither mode guarantees perfect recognition.
Here is an illustration, not a tested model result:
You say: “Set the timeout to ten—no, thirty seconds. Only for the upload route.”
A useful cleaned-up draft would be: “Set the upload-route timeout to 30 seconds.”
The number matters. So does “only.” If either changes, the agent gets a different job.
What would it cost?
Google’s current API pricing estimates about $0.009 per minute for live transcription and $0.005 per minute for recorded audio.
That is roughly 54 cents or 30 cents per hour, respectively. These are estimates using Google’s assumed audio and text token volumes—not fixed per-hour fees. Coding-agent work, hosting, and other services cost extra.
The pricing page also marks free-tier content as usable for product improvement, unlike the paid tier. Use non-sensitive test recordings unless your organization has approved the service and terms.
Why Maestro users care
A possible workflow is to dictate a bug report, check the resulting task, then give it to your coding agent. That could reduce typing without removing the review step.
Do: Try a few recordings containing real file names, numbers, and corrections. Read the final text before submitting it.
Do not: Connect polished transcription straight to deploy, delete, or permission-changing actions.
Why: Speech recognition and text cleanup can both change meaning.
For a live integration, note the documented ten-minute session limit. Live mode does not provide speaker labels or word-level timestamps; those require the recorded-audio route.
Bottom line: Let it clean up your sentences. Keep the decision about what to do next.
Maestro’s opinions and summaries are AI-generated. This brief received independent AI editorial review.