When dictation drops words, most people blame the software and switch apps — Apple's dictation to Wispr Flow to something else — and the misses follow them. That's the tell. Modern speech models are extraordinarily good at transcribing clean audio. What they can't do is transcribe audio that isn't there.
What the model actually receives
Speech recognition doesn't hear the way you do. Your brain uses context, lip memory, and attention to reconstruct a word that got quiet; a model gets a waveform, and if a word's energy dropped 10 dB because you turned your head mid-sentence, the model receives something closer to a murmur under fan noise. It does the statistically sensible thing: skips it, or guesses.
The three audio failures behind most "dictation is bad" complaints:
1. Level swings from movement. Dictation is where you move most — pacing while thinking, turning to read a reference document, leaning back mid-thought. Every movement changes your distance and angle to the one mic that's listening, and consonants at the ends of trailing sentences are the quietest, most information-dense parts. They vanish first — which is why your last few words per sentence miss most often.
2. Room sound smearing consonants. A mic that's too far away captures your voice plus its reflections a few milliseconds behind. Human listeners tune the smear out; to a model, "sit" versus "sits" lives in exactly the crisp high-frequency detail that reverb blurs. If your transcripts drop plurals and "-ed" endings, distance is the suspect — placement fixes half of it for free.
3. Aggressive processing. Noise suppression built for calls happily eats quiet speech onsets, and "voice isolation" modes can gate off your first syllable after a pause. For dictation, gentler is better: models handle steady background noise far better than they handle chopped speech.
Test it on your own transcripts
Dictate one paragraph twice: once frozen at your ideal mic distance, once moving the way you naturally do. Compare miss rates. Nearly everyone finds the frozen version near-perfect — which localizes the problem precisely: it isn't the model, it's coverage of a moving speaker.
Fixing coverage instead of freezing yourself
You can hold still forever, or make the audio follow you. This use case is why SoundMesh exists: several mics covering desk, second monitor, and pacing range; the app continuously hands your voice to whichever mic hears it best; your dictation tool sees one steady virtual microphone with far more even levels than any single mic can deliver from a moving speaker. One mic live at a time — switching, not summing — so there's no phasey artifact for the model to fight either (why summing backfires).
Whatever dictation app you've bet on, it's already good. Feed it a signal that survives the way you actually work, and the accuracy you were promised shows up. Run the test above with SoundMesh on your own setup — your transcripts are the only benchmark that matters.