It's 11pm, and you're deep in a design doc for the payment settlement refactor. You've been explaining the schema change for three minutes, building momentum, when the dictation stops accepting words. The message at the top of your dictation app reads: "Free tier: 500 words this month. Upgrade to continue." You're 347 words into a 1200-word spec. You have two choices: stop dictating, switch to typing (and lose the thinking voice you had), or pay fourteen dollars a month for unlimited words.
Why the Bottleneck Moved
Code is no longer what's slow. You write intent now, not syntax. Cursor autocompletes the implementation; you spend the time describing what it should build, why it should build it, and what failure looks like. A design doc that used to take typing is now thirty minutes of explanation, because you're not writing instructions for yourself, you're writing context for an LLM to read and reason about.
The architects and technical leads who started using voice to speed up that intent-writing discovered a hard constraint: cloud-based dictation charges per word, caps the free tier aggressively, and treats "free" as a trial designed to expire. Wispr caps free at 1500 words a month. Willow's free tier has a similar wall. Superwhisper locks you into one particular IDE. They're building around the old workflow, not the new one.
The Local Moment
Marcus, a backend engineer at a fintech, tried three different voice tools before settling on one that didn't charge per word. The first two worked fine until they didn't: they'd run out of monthly quota mid-explanation, force a login screen, or suggest an upgrade the moment a tool became useful. He refuses cloud transcription anyway; his design docs reference payment schemas, incident patterns, and customer query structures. Sending that audio to the cloud, even encrypted, violates his instinct about code IP.
Recitey's technical approach removes the problem entirely. It runs Whisper, the speech-recognition model from OpenAI, locally on his device. No audio leaves the computer. No word counter. No monthly limit. The bottleneck is no longer whether he can afford to dictate; it's whether he has things worth dictating.
What Actually Changes
When the dictation barrier falls, the writing changes. Marcus noticed it within a week: he stopped pre-planning what he'd say, because stopping to pre-plan is just giving up the voice. He spoke through entire sections in one breath, captured his thinking as it shaped, and then let the local model transcribe it into text. The prose came out rough, exactly as rough as speech sounds when you write it down verbatim, but the thinking was intact.
That's where the second part comes in. The Pro tier runs a rewrite pass on the rough transcription, polishing the dictated speech into clean prose in about two seconds. For Marcus, the decision to upgrade was straightforward: the free tier solved the IP concern and the quota problem; the Pro tier solved the polish problem. He could use free alone and accept rougher prose. He chose to upgrade because the trade-off made sense.
The Actual Limitation
Local dictation has one real limitation, and it's worth naming honestly. Whisper-large-v3 achieves 96.3% word accuracy on LibriSpeech's test set, which is exceptional, but "exceptional accuracy" is not "zero errors." Mishearings still happen. Words get confused. Accented speech or technical jargon sometimes loses fidelity. Cloud-based systems can use your correction history and your voice profile to learn and improve over time. Local processing can't. You get accuracy out of the box, but not personalization that compounds over time.
Marcus accepted this trade-off. He edits the occasional mishearing ("settlemint" becomes "settlement") in exchange for not paying per-word fees and not uploading code audio to the cloud.
Who Wins With This Approach
The local-first model appeals to a specific segment: engineers who own a new workflow shape (speaking intent for LLM collaboration) and care about either privacy (code IP) or consistency (no monthly surprises, no cap that hits mid-thought).
It doesn't appeal to everyone. Someone using voice to transcribe interviews or lectures might prefer cloud-based because the post-processing and speaker identification is better. Someone on a Mac will find better native options. Someone who values convenience over privacy will find it easy to stomach a cloud upload.
But for developers building with LLMs, describing technical decisions out loud, and refusing to let a subscription meter sit between them and their thinking, local Whisper with no cap is a clarity it's hard to go back from.