Marcus is a backend engineer at a fintech. He used to measure typing speed in lines-per-minute. Now he measures it in prompt quality per sitting. He's not writing code; he's writing intentions for Cursor to fill in. The interface hasn't changed. The work has.
Three months ago, he tried Wispr Flow's free tier. It capped at 2,000 words per month. He hit the limit halfway through a design doc on Thursday. By Friday morning, the half-finished thought had cold. He switched back to the keyboard. The momentum was gone.
Most voice tools miss the gap: how fast intent flows when you speak versus how slowly it dies when you stop.
The Work Changed, the Tools Didn't
Developers used to think of dictation as a way to avoid RSI. Dictate code comments faster. Reduce wrist strain. It's a nice-to-have, not a necessity. That framing was built for a different era, when developers actually typed code line by line.
LLM-assisted development inverted it. You're not writing code anymore. You're writing prompts that describe what the code should do. You're writing PR descriptions that justify why you changed it. You're writing Slack threads that walk through a bug investigation. The amount of prose you generate per day has doubled or tripled, while the amount of actual code-typing has dropped.
If you're pairing with Claude or GitHub Copilot, you spend more time writing context and specification than you do writing syntax. That context is where voice wins. You can think it out loud and let the tool catch it. No stopping to craft the perfect sentence. No switching tabs to paste a stack trace. Just flow.
But flow stops when you hit a word cap.
Why Cloud Dictation Breaks for Developers
The problem isn't the tool. It's the business model. Cloud-based transcription services charge by API call. Wispr charges $14 a month for unlimited, or $0 on free with a 2,000-word ceiling. Willow does $12 a month, capped at 3,000 words free. Superwhisper is $8.49, capped at 1,000 words free.
Each one has made a trade-off: cap the free tier to funnel users into paid plans. It's a sensible business model. It's also why Marcus stopped using it.
The second problem is opaque. When you send audio to the cloud for transcription, you don't know what happens to it. Does the service log it? Retrain on it? Sell anonymized transcripts? For a fintech engineer writing code and design docs with internal infrastructure details, that ambiguity isn't acceptable. Cloud is too far from the data.
Local-first speech-to-text flips both constraints: no word cap because there's no per-word API cost, and no data leaving the device because the model runs on your machine.
Local Whisper Changes the Equation
Whisper, OpenAI's open-source speech recognition model, hits 96.3% word accuracy on LibriSpeech. It runs locally on your device. There's no cloud call, no API cost, no rate limiting. You can dictate 10,000 words in a single sitting and hit exactly zero metering.
Recitey packages Whisper as the free tier. No word counter. No monthly cap. No upsell friction when you're mid-thought. The architecture does the heavy lifting: speech runs locally; the rewrite happens in the cloud only if you want it polished into publication-ready prose.
That architectural choice, local dictation plus optional cloud rewrite, means the free tier is genuinely unlimited. It's not a trial. It's not crippled. It's a complete speech-to-text tool with one job: get your voice out of your head and into the document without losing the thread.
How Marcus Uses It Now
It's 11pm, and Marcus is writing a design doc for payment settlement logic. The idea is complex. It involves ledger writes, idempotency keys, reconciliation timing. None of it fits a two-sentence email.
He opens Cursor. He hits the voice button in Recitey. He explains the approach out loud: how the state machine flows, why certain operations need to block, where the race condition lives. Fifteen minutes of talking. No stops. No rewrites. Just thinking aloud while Recitey catches every word.
When he's done, he has a full paragraph of raw speech-to-text. It's rough. It repeats itself. But it's complete. The thought is there. He can edit. He can publish. The momentum held.
Tab-complete in Cursor catches typos and awkward phrases as he edits, which saves another pass. Raw voice dump followed by Cursor polish is faster than typing from scratch.
Three months ago, cloud dictation would have cut him off mid-paragraph with a word warning. Instead, he typed the rest on the keyboard, lost the thread, and finished the doc the next morning. Now the doc is done the same night.
The Trade-Offs Are Real
Local Whisper is fast and uncapped, but it's not perfect. It will mishear technical terms. It will add disfluencies (um, uh) that the cloud version would clean. If you need publication-ready output on the first pass, local alone is rough.
That's why Recitey makes the rewrite optional, not mandatory. If you dictate a PR description in your local Whisper, you get clean speech-to-text instantly. If you want it polished into something more formal or grammatically precise, you send it to the cloud rewrite. That's where Pro comes in.
But the free tier is complete. You're not using Wispr or Willow and hitting a wall. You're using a real tool that knows when to stop gatekeeping.
This Matters Because the Bottleneck Shifted
For the last ten years, voice dictation was an accessibility feature or an RSI workaround. Nice to have. Now, with LLM-assisted development, it's a workflow cornerstone. Your ability to think aloud without artificial stops has become competitive.
Most voice tools still price like they're protecting a speech-to-text API call. They're not wrong; transcription costs money at cloud scale. But Recitey's local-first model means there's no per-word cost. The bottleneck isn't transcription anymore. It's clarity. It's knowing whether your polished thought is actually thought, or just smooth noise.
That's a different game.