You're designing a payment flow at 11pm, speaking directly into your voice doc. Then the word counter hits. Most free voice tools cap the free tier because cloud transcription costs them money. But that cap kills the thinking. The architecture was clear in your head; now it's half-written, and you've switched back to typing by morning.
The workflow that changed
You didn't always think in voice. Three years ago, you typed code in silence. The bottleneck was fingers. Now, with Claude, Cursor, Copilot, the bottleneck's explaining your intent. Models need specs, not syntax. You're spending less time typing implementation and more time articulating the problem. The edge case, the sequence, the why.
Voice's natural for that articulation. Your mouth's faster than your hands for long-form thinking. But most voice tools weren't built for transcription. They're replacing typing for everyday chat. They meter you. They assume you'll dictate your messages in bursts, not sustain a 10-minute design monologue.
Hitting a word cap mid-monologue isn't like hitting a character limit mid-email. It's a brake on thinking.
The local-first trade-off
You don't send code over the internet unless you've got to. IP concerns are real. Fintech codebases have regulatory shadow. The backend at your company stays on-device in IDE and local tools. Cloud dictation services cache audio. They promise it gets deleted, but you're not shipping unencrypted payment logic over the wire to prove a point.
Local speech-to-text exists. OpenAI's Whisper-large-v3 hits 96.3% accuracy on LibriSpeech and runs on your machine. No servers. No upload. The model's 3GB. It matters because it means zero variable cost and zero data leak.
Recitey's free tier runs Whisper locally. No word limit. No metering. No cap. You can design the full spec at 2am if you need to. The tool's there when you think.
Why the pricing story is backwards
Wispr Flow charges $14 a month. Superwhisper charges $8.49, locked behind indie distribution. Both cap their free tiers because the cloud infrastructure costs them money per transcription second. The premium tier pays for the rewrite engine, the polish that turns your rough draft into a clean sentence.
Recitey inverts that. The dictation's free because it runs local, zero per-use cost. The rewrite is where the LLM polish happens. Your rough "settlement retries if network fails" becomes "if the network times out, the system retries settlement within 30 seconds". That's the cloud call, where the variable cost sits.
It's a different business model because it's a different workflow assumption. The tool doesn't meter the transcript. It meters the transform.
The Cursor era of voice
You switched from VS Code to Cursor six months ago. Not for the editor; for the tab-complete. Fewer keystrokes between voice input and clean output. It's a small friction that compounds across a week of design docs.
Recitey works the same way across Windows. Slack, email, browsers, Linear, Notion, Cursor. The voice input flows to the system clipboard. The rewrite's optional, not mandatory. You can't buy into a single IDE or agent ecosystem. You'll voice-draft in whatever context you're already thinking in.
Marcus's 11pm design doc wasn't interrupted by a technical limit; it was interrupted by a pricing boundary with nothing to do with the work.