Speech-to-Text
No more "um," no more "uh,"
they're gone before you see them.
Automatic meeting transcripts have always been drafts — good enough to summarize from, not to read as-is. Google DeepMind's new speech model, Gemini 3.5 Transcribe, supports 85 languages and strips filler words in real time as it transcribes. Here's how it works, and where verbatim records still need the old, messier approach.
Google just played a new card
in the meeting-notes war
Automatic meeting transcription is already a crowded field. Beyond dedicated startups like Otter.ai and Fireflies.ai, Zoom AI Companion and Microsoft Teams' Copilot ship the feature standard, and Google itself has kept improving Google Meet's captions and summaries. But most of these tools still work in two steps: transcribe everything verbatim, fillers and all, then hand that raw transcript to a summarization AI to clean up.
Google DeepMind's new release changes that premise. Gemini 3.5 Transcribe supports 85 languages and detects filler words as they're spoken, producing already-cleaned text in real time instead of raw text to be fixed later. That's the real shift — going from "clean it up afterward" to "output it clean from the start." Google's own technical writeup says the filler-detection layer combines speech timing with acoustic features in a dedicated classifier model.
The bet isn't language coverage —
it's removing fillers on the spot
By raw language count alone, Gemini 3.5 Transcribe isn't actually the widest-reaching model out there. The real differentiator is elsewhere.
On raw language coverage, OpenAI's publicly documented Whisper large-v3 actually covers roughly 99 languages — more than Gemini 3.5 Transcribe's 85. So the differentiation isn't the language count; it's the timing of the cleanup. Most existing tools remove fillers in post-processing, after the full transcript is finalized. Gemini 3.5 Transcribe folds that removal into the streaming pass itself, so the text that actually appears on screen is filler-free from the moment it's rendered.
That distinction sounds small but matters in practice. The old workflow was: transcribe with fillers intact → hand off to a summarizer → review the summary. With Gemini 3.5 Transcribe, the raw transcription output is already close to a usable meeting-notes draft. One entire cleanup step, upstream of any summarizer, effectively disappears.
Who benefits, and how
The size of the payoff depends a lot on your role. Three angles worth separating.
Meeting owners and PMs
The "make it readable" step of writing up notes largely disappears. Teams with a high meeting cadence see the biggest cut in monthly cleanup time.
Teams with lots of non-native speakers
85-language support plus automatic filler removal is a bigger win for international teams where English (or any shared language) isn't everyone's first language — non-native speakers tend to hesitate more, so the cleanup is more noticeable.
Legal and medical settings needing verbatim records
This one calls for caution. In negotiation records or counseling notes, hesitation itself can be part of the evidentiary or clinical record — automatic filler removal can conflict with a verbatim requirement.
Not cleaned up afterward —
output clean from the start.
What happens next
The thing to watch is when this transcription engine gets folded into Google Meet's captions and notes, or into Google Workspace's summary features. As a standalone API, it stays a tool for developers and early adopters; the moment it lands as a default in Meet, usage jumps by an order of magnitude. Worth watching the Google Workspace Updates blog for a rollout timeline.
Three practical next steps: (1) pilot it in internal meetings and check how it interacts with your existing summarization AI; (2) if you're in legal or medical work, confirm with the vendor whether automatic removal can be toggled off; (3) run an early trial in international meetings with non-native speakers and measure the actual cut in cleanup time.
Is the "missing hesitation"
really unneeded information?
Removing fillers improves readability, but something is lost in the process. Hesitations and pauses often carry real non-verbal signal — a speaker's uncertainty, emphasis, or hedging — and in negotiation records or clinical notes, that "pause" itself can be meaningful context. If automatic removal is on by default, users may lose important nuance without ever noticing.
There's also the question of the filler-detection model's own accuracy. Depending on dialect or speech patterns, it's plausible that a genuine word could be misclassified as a filler and dropped. Google hasn't published a specific misdetection rate yet, so how stable the accuracy stays in real-world use remains untested for now. For anything requiring verbatim fidelity, running the raw transcript alongside the cleaned one — rather than turning off the raw record entirely — is the more realistic setup today.