Google's New Transcription Model Is Genuinely Impressive. Here's the Trade-Off.

Google launched Gemini Transcribe this week, and it's worth taking seriously rather than dismissing. The company says its new speech-to-text pipeline is around 70% faster end-to-end than its previous model, with a live-speech error rate down to about 5.5%, support for 85 languages out of the box, and — the part that actually matters day to day — it cleans up filler words and false starts as it transcribes, instead of faithfully typing out every "um" and "no, the other one."

That last part is a real usability win. Raw verbatim transcripts of normal speech are messy, and cleaning them up by hand is most of what makes transcription tedious. If Google's claims hold up, this closes a real gap that on-device models — including the one behind OffgridScribe — mostly haven't solved yet.

So what's the catch?

Gemini Transcribe is a cloud model. Your audio has to leave your device and reach Google's servers before you get text back — there's no on-device version of this being described anywhere in the announcement. That's a completely reasonable trade for a lot of use cases. It's a different trade for others: a journalist protecting a source, a lawyer with privileged conversations, a therapist's session notes, or just someone who'd rather a tech company not have a searchable transcript of their personal conversations sitting on a server somewhere.

To be fair to Google, their enterprise products generally offer stronger data controls — limited retention windows, options to avoid using data for model training, compliance certifications for regulated industries. The free, consumer-facing tier of Google's AI products has historically had looser defaults than the enterprise tier, with data sometimes eligible for human review and retained for extended periods unless you actively opt out. It's worth reading the specific terms for whichever tier you'd actually be using before assuming "Google" means one privacy posture across the board.

Two different products, not a ranking

We're not going to pretend OffgridScribe matches a frontier lab's latest cloud model on raw transcription quality — that's not a fair fight, and claiming otherwise wouldn't survive five minutes of anyone actually testing both. What we'd say instead: the two things solve different problems. Gemini Transcribe is for when you want the best possible transcript and don't mind the audio leaving your device to get it. OffgridScribe is for when "leaving your device" is the part you're not willing to accept, even if it costs you some polish.

Neither answer is wrong. They're just answers to different questions, and it's worth being honest about which question you're actually asking before you pick a tool.

More from Offgrid Studio

OffgridScribe transcribes recordings using WhisperKit entirely on your iPhone — no upload, no account, no cloud model in the loop. If it's noise cleanup rather than transcription you need, OffgridVox works the same way.