The Gemini 3.5 Transcribe announcement got me thinking, and not jus... - Vibeus
The Gemini 3.5 Transcribe announcement got me thinking, and not just about accuracy scores.
Here's the thing about speech-to-text: accuracy is the headline, but it's never the whole story. What actually determines whether someone can use these tools day to day is what happens *around* the transcription.
A few moments worth designing for:
**The moment of uncertainty.** When someone says an unfamiliar name, does the transcript show a confident-looking wrong word, or does it signal "I'm not sure here?" A subtle underline, a highlighted guess — anything that says "check me" instead of silently committing to an error. Interfaces that never admit doubt force people to proofread every line as if it might lie to them. That's exhausting.
**The moment of correction.** Fixing a misheard word should take one tap, not a detour through selecting text, opening a keyboard, and retyping. And corrections should ideally teach the system — my aunt's name shouldn't be mangled fresh every single morning.
**The moment of recovery.** This one matters most and gets discussed least. If you have an accent, a speech difference, or you simply talk fast when excited, a transcript full of errors can feel like the interface telling you to speak "properly." Good design assumes the mismatch is the system's problem, not the speaker's. A gentle "want to teach me this word?" lands very differently than a wall of confident nonsense.
Not everyone has identical hands, identical accents, or identical attention — a fact some interfaces still find genuinely surprising.
If you're building with transcription somewhere, here's a small thing to try this week: deliberately say a name the model won't know, then watch how the product behaves. Does it flag uncertainty? Can you fix it in one move? Does the fix stick?
Those three answers tell you more about accessibility than any accuracy benchmark.