Marcus is 11 hours into a payment settlement redesign when the thinking finally crystallizes. He's been working through architecture tradeoffs, modeling the data flow, running scenarios against the settlement engine. The spec is clear in his head. He reaches for his phone, opens his voice-to-text app, and starts dictating. For long-form exploratory work before code, voice is faster than typing. It's how he thinks out loud, and the tool transcribes it into a design doc.
Twenty minutes in, mid-paragraph, the recording stops. Word limit hit. The tool won't listen anymore. The thinking breaks. He switches to typing the rest, hurriedly, fragmented, full of the mental shortcuts that made sense when speaking but don't read right on the page. The next morning, he'll have to rewrite it. For now, the design is half-narrated and half-typed, and neither captures the actual thinking.
This isn't a rare moment. It's his entire workflow.
The Work Changed, the Tools Didn't
Developers' relationship to writing has fundamentally shifted in the past eighteen months. The bottleneck used to be typing speed. Now it's clarity of intent. When you work with LLMs, you don't write code. You explain what you want the code to do. The model writes it. The work is in the prompts, the specs, the long-form explanations.
Voice is measurably faster for this kind of work. A 1200-word design doc takes forty minutes to type and eighteen minutes to dictate. The speed advantage isn't just about words per minute. It's that you're not stopping to arrange thoughts into formal written English. You're narrating them. The model reads the narration and fills in the gaps. You explain payment settlement logic at the speed you think it, not at the speed your fingers move.
But every consumer voice tool imposes a cost at scale. Wispr Flow's free tier: 600 words per month. Dragon NaturallySpeaking: 500 words per recording. Superwhisper: $8.49 per month for unlimited access. They frame these as service tiers. What they actually are is rationing designed to force conversion.
For a developer dictating three design docs a week, an incident postmortem, three Slack threads explaining bugs, and half a dozen PR comments, all running 800 to 1500 words each, these caps evaporate in four days. You hit the limit. You stop dictating. You go back to typing, which is slower, which breaks your flow, which means the work takes longer and the thinking is less clear.
Local Whisper Breaks the Metering Model
Whisper is an open-source speech-to-text model released by OpenAI in September 2022. It runs locally on your device. On a 2024 laptop, it's faster than sending audio to a server and waiting for transcription to come back. Latency is lower. Data never leaves your machine. Cost per word is exactly zero.
And yet: nearly every voice tool uses Whisper but adds a metering layer on top anyway. Otter.ai runs equivalent models in the cloud and caps the free tier at 600 minutes per month. Wispr runs locally but caps at 600 words. They're not metering because transcription is expensive. Whisper proves it isn't. They're metering because cloud infrastructure and billing systems are standard industry practice, and free tiers with no limit would break the conversion funnel.
Recitey runs Whisper locally on your device. No word cap. No session limit. No monthly meter. The free tier is the full speech-to-text engine. The Pro tier is the rewrite layer, the refinement that polishes rough voice into structured prose in two seconds. You're solving two different problems: transcription and refinement. You should pay for the problem you're solving, not for both.
This changes what "free" actually means. It's not a degraded tier designed to frustrate you into conversion. It's a full tool with no artificial limit, paired with an optional premium service if you want the output refined automatically. The metering disappears because the economics change when software runs on your device instead of on a server.
Data Stays On Your Machine
Marcus refused cloud dictation for six months because of one reality: code in design docs. You write a design document that mentions code examples. You read code snippets aloud. You walk through a database schema verbally. A cloud service records all of it. You can request deletion, set privacy flags, read the terms of service. But the architecture remains: audio leaves your device, lives on a server, gets transcribed, maybe gets retained for machine learning, maybe gets deleted after thirty days.
For payment settlement code, for authentication logic, for proprietary IP, that's not a trade-off you make consciously. It's a risk you accept because the better tool isn't available.
He also switched to Cursor instead of VS Code for one reason: tab-completion that feels native to your thinking. In Cursor, you type a few characters and the model completes the line. The flow stays uninterrupted. The thinking doesn't break for context switching. The same principle applies to dictation. Every sync to a server, every privacy dialog, every knowledge that audio is leaving your device, it interrupts the thinking.
Recitey's local model means your code stays on your machine. Transcription happens on your device. The only data that leaves is the final text you explicitly choose to send somewhere. Your design docs, incident postmortems, code examples, they stay private by architecture, not by policy. For payment systems, for security-sensitive code, for proprietary data, it's the difference between acceptable risk and no risk at all.
The Counter Is Gone
Most voice tools display a word counter as you speak. "You have 67 words left before hitting the cap." It's meant to set expectations. Instead, it becomes a ceiling. You're not thinking about what to say next. You're thinking about how many words you have left. You're calculating whether to finish this thought or save words for the next section.
Recitey has no counter because there's no cap. This sounds minor. It's actually the core insight. The absence of a meter is the presence of freedom. You can finish the thought. You can explore a tangent. You can go into depth without watching the clock.
When you design at 11pm, you're not optimizing for brevity. You're trying to get the thinking out before it evaporates. The tool should enable that, not constrain it. Removing the counter removes the friction entirely.
Who This Changes
The backend engineer who redesigns critical systems at 11pm, after the synchronous work is done. The on-call responder documenting an incident at 2am, thinking clearly but needing to move fast. The indie builder in Cursor building features for clients and explaining architecture to stakeholders. The Series B architect who dictates PRDs before writing the actual requirements document.
These aren't edge cases or premium-user moments. They're the shape of development in 2026. Long-form explanation is the work. The tool should reflect that, not fight against it.