You're at 11pm writing a design doc for an async review in the morning. Speech is faster than typing, you're explaining the data flow, the edge cases, the decision tree. Then the cloud transcription tool cuts you off. Word cap. You close the app and finish typing it out, fragmented and tired. By morning, you're rewriting what should have been fluid prose.
The design-doc voice problem
Cloud dictation tools were built for transcription, not long-form technical writing. Developers hit word caps mid-flow. The rhythm breaks. The tool becomes friction instead of aid. Most developers I've talked to abandon voice transcription after one week because of it.
The free tiers all cap you. Wispr Flow ($14/month) gives you 2,000 words on free and cuts you off. Willow ($12) has a 500-word monthly limit. Superwhisper ($8.49 one-time) is uncapped, but it's macOS only. There's a reason the pricing model is tight: cloud transcription costs money per API call. Every word you dictate is a compute charge on the vendor's side. So they meter you.
The local-first structure
Recitey runs Whisper, specifically Whisper-large-v3, which hits 96.3% word accuracy on LibriSpeech, on your device, offline. No word counter. No cap. No monthly bill for transcription. The model is 140MB. It runs once and you own the output. The pricing (Recitey Pro is for the rewrite layer, the polish step) isn't metering your speech. It's metering the edit intelligence, which is a different cost structure entirely.
This matters if you're thinking about code IP. Marcus, a backend engineer building payment settlement logic at a Series B fintech, won't use cloud dictation. He's pasting implementation notes and SQL schema snippets into design docs. A third-party transcription API is a data leak risk, even if the vendor claims encryption, you're still trusting someone else's infrastructure. Local means you control the boundary.
The intent-first workflow
The job of dictation in a developer's day has changed. You're not transcribing meetings anymore. You're explaining intent to an LLM. You're drafting a Cursor prompt for a feature. You're writing a GitHub PR description that's detailed enough that a reviewer doesn't have to ask clarifying questions. That's long-form, specific, technical speech. A 500-word limit stops you cold.
When dictation doesn't interrupt you, the flow continues. You've stopped context-switching between voice and hands. Cursor's tab-complete can polish the rough output in one keystroke. No cloud round-trip latency. No metering anxiety. The tool vanishes.
Trade-offs you're actually making
Local Whisper isn't perfect at code. It'll transcribe const correctly but sometimes stumble on library names or variable-naming conventions if they're unusual. You're trading transcription perfection for zero friction and zero cost. Most developers who've stayed with local dictation tell me it's worth it because they're not abandoning the tool mid-document anymore.
The uncapped free tier isn't a forever promise. It's a structural choice: if the bottleneck is your speech, not cloud infrastructure, we're not metering your speech. If you want the cloud-side rewrite that polishes your rough draft into publication-ready copy in under 2 seconds, that's Recitey Pro.
Voice writing fails when the tool is optimized for the transcription business, not for the developer's actual workflow. The word counter isn't a feature. It's a confession that the tool isn't built for what you actually need it for.