Design docs, architecture decisions, incident postmortems, PR description context; these are what take time now. Not typing code. Voice is faster for the long explanations that tell a model what to build.
Why developers are already voice-writing, but hating it
Marcus, a backend engineer at a fintech in Stockholm, switched to Cursor specifically because tab-complete halves the number of rewrites he gets from Claude on each iteration. Every keystroke now carries more intent, less noise. So when he sat down at 11pm to draft a design doc on payment settlement logic, voice felt obvious: he could explain the architecture faster than typing it, with Cursor's autocomplete catching the half-thoughts and cleaning them up.
Twenty minutes in, he hit a wall. Wispr Flow's free tier has a 2000-word cap. He was 1800 words in. The cap interrupted him mid-explanation of the state machine. He lost the thread. He finished the doc the next morning, fragmented, less coherent.
Cloud dictation services lock in word caps to force an upgrade. The economics are straightforward: they bill per minute or per word, so the marginal cost of letting you dictate forever is real. They've got cloud infrastructure to pay for. But Marcus doesn't care about their economics. He cares that he was thinking clearly, the tool interrupted him, and now the doc's worse.
The IP problem nobody talks about
Local-first infrastructure changed something else for Marcus: he stopped using any cloud transcription for code-related work. Payment settlement logic, database schemas, API structures; he doesn't trust sending that audio upstream to Wispr or Otter. The terms of service probably disclaim liability, and the data's probably deleted, but "probably" isn't a guarantee when code IP's at stake.
So Marcus was trapped: voice is faster for explaining systems, but the tools that do voice transcription either cap you (free tier) or require cloud upload (privacy risk). Neither fits the workflow.
The moment uncapped local Whisper changes the shape of work
Recitey runs Whisper locally on his Windows machine. The transcription never leaves the device. The free tier's got no word cap; he can dictate an entire design doc without interruption. The local model (Whisper-large-v3 achieves 96.3% word accuracy) is accurate enough that Cursor's tab-complete can clean up the remaining 3.7% of rough edges in seconds.
For Marcus, this changes the day-to-day:
- 11pm design docs: He can explain the state machine completely, in one voice session, without watching a word counter. Flow doesn't break.
- Incident postmortems: Long explanations of what went wrong, what the decision tree was, what the fix cost; all captured without uploading code context to the cloud.
- PR descriptions with context: He can explain why he chose a particular pattern instead of typing a generic summary.
The bottleneck shifts. Accuracy matters less than interruption. Interruption kills thinking.
Where voice isn't a fit (and probably never will be)
Code review comments on specific lines? Voice is slower. You've got to describe line numbers, context, why the change matters. Typing's clearer. Same with commit messages that need to be terse and scannable. Voice doesn't compete there.
But for thinking-intensive work, such as architecture decisions, investigation reports, or design rationale, voice is faster and clearer than typing, provided the tool doesn't interrupt you.
What the pricing structure actually says
Wispr Flow charges $14/month for unlimited. Superwhisper is $8.49/month indie. Willow runs $12/month. All cap their free tiers to force an upgrade. Recitey's free tier runs Whisper locally, uncapped. The reason's simple: running Whisper on your device costs Recitey nothing. Every transcription is free to them. The cap isn't a technical limitation; it's a pricing lever.
When a tool removes the cap, it's betting that you'll eventually want the cloud rewrite feature (which does cost them; the language model that cleans up the dictation runs on their servers). But you can ship docs, postmortems, and design decisions on local Whisper alone. No upgrade required.
Marcus doesn't upgrade for uncapped dictation because he doesn't need to. He upgrades for the rewrite if the raw transcription isn't clean enough. The pricing aligns with the actual cost structure, not with artificial scarcity.
The frame shift: flow over accuracy
For a generation of engineers, the bottleneck was output speed: type faster, ship faster. That was true. Mechanical keyboard culture made sense.
The bottleneck now is input clarity: explain your intent clearly enough that the model can build it right. Voice is faster for intent. Typing is faster for syntax. The tools that recognize this and get out of the way (no caps, no cloud upload for code, model accuracy that's "good enough" for intent) will be the ones that fit the new shape of work.
The word counter isn't the problem. The interruption is.