You're writing a design doc at 11pm. The thought is clear. You're speaking fluently through a complex workflow decision, and then your voice tool stops. Word limit hit. You switch to typing to finish. Your flow breaks. What you get is two documents stitched together: the confident voice section and the fragmented typed section below it. It's 11:47pm and you're supposed to hand this off in the morning.
The Bottleneck Shifted
A few years ago, the developer bottleneck was typing speed. "Just code faster," the advice went. Fast typists shipped faster. That was the frame.
Now it's different. You're not bottlenecked by typing speed anymore. You're bottlenecked by articulating intent clearly enough that a model understands what you want to build.
That shift happened quietly. It started with GitHub Copilot making auto-complete smarter. Then Claude and GPT-4 started handling complex architecture questions. Now LLMs are in your PR reviews, your design docs, your incident postmortems. The work you do at the keyboard has become the work of describing what you want the model to do.
That means more words per day. Longer design docs. Clearer specs. More detailed PR descriptions. Your Slack explanations of a bug investigation are now mini engineering documents because the reader might hand them off to a model. An incident postmortem that used to be two paragraphs is now six paragraphs because you're documenting not just what happened, but the architectural implications.
A voice tool that meters your dictation by the word doesn't fit this workflow anymore. It's like having a Slack message limit enforced mid-thread. It interrupts thinking at exactly the wrong moment.
Marcus at 11pm in Cursor
Marcus is a backend engineer at a Series B fintech in Stockholm. He works on payment settlement systems; the infrastructure that ensures money actually moves when it's supposed to. His tool stack: Cursor, Linear, Notion for design docs, GitHub for PRs, Sentry and Datadog for incident response.
He chose Cursor over VS Code specifically because Cursor's tab-complete reduces the number of times he has to rewrite voice drafts. When you're dictating code-adjacent content, every rewrite breaks your flow. Tab-complete keeps him in flow. He's refused to switch back to VS Code even though his team uses it, because the voice-writing penalty is too high.
At 11pm on a Tuesday, he's writing a design doc about payment settlement retry logic. Not a blog post. Not a spec. A design doc (the kind that lives in Notion and gets read by the team before the sprint starts). He's speaking through the architecture. Edge cases. Data schema. Retry windows. Idempotency keys. The complexity is high and he's explaining it clearly because he's thinking out loud. The doc is flowing.
Cursor keeps him in flow. Then Wispr Flow (his current voice tool, $14/month for the subscription) stops recording. Word cap hit. He's at 847 words; Wispr's free tier caps you at 1,000 words per month. He should've switched tools on day one, but it's 11:47pm and the thought isn't finished.
He switches to typing. The prose changes immediately. His typed sentences are shorter, choppier. The voice section flows. The typed section doesn't. When he reads it back in the morning, it's obvious where the voice ended and the keyboard started. He spends 20 minutes rewriting it to make it cohere.
He's also stopped using cloud transcription for code examples. Payment settlement code in someone else's cloud? That felt like a risk he didn't need to take. Compliance, IP, the usual concerns. If cloud transcription services got breached, it's not just his architecture that's exposed; it's financial infrastructure. So he types code snippets when necessary.
Local Whisper Changes the Economics
Here's what changed for him: Whisper runs locally on your device. Not in the cloud. That structural difference cascades.
First, code stays on your machine. No upload. No question about data residency or compliance. Your design doc about the retry logic doesn't leave your disk until you decide to share it. Fintech compliance teams get less nervous.
Second, there's no per-word cost, so there's no word limit to enforce. Wispr Flow's free tier caps you at 1,000 words per month. Willow caps you at 2,000 words per month. Both are cloud-based tools, which is why they meter; cloud transcription has variable compute costs. Recitey runs Whisper locally with zero variable cost per word, which means zero reason to meter you. Write as much as you need. The free tier is uncapped.
Whisper is the open-source speech recognition model from OpenAI. It's trained on 680,000 hours of multilingual audio and hits 96.3% word accuracy on LibriSpeech benchmarks. It's not perfect, but it's good enough for design docs, specs, and PR descriptions. The important part: it runs on your device, not Anthropic's servers or OpenAI's infrastructure.
The technical tradeoff is real. Whisper runs on first transcription, which is slightly slower than streaming cloud dictation. First transcription might take 2 to 5 seconds depending on the length of what you just said. Cloud dictation streams in real-time, which feels snappier. But real-time streaming is also why Wispr and Willow have to meter; they're paying for cloud infrastructure per second.
But the upside is clearer: unlimited words, offline capability, code privacy, and no mid-thought cutoff. Marcus finishes the design doc. The prose is unified. He doesn't have to rewrite half of it in the morning before he ships it.
What Doesn't Change
Free tier is still free. There's no word counter watching you. There's no cloud upload of your design docs or incident postmortems.
What does cost money is the rewrite engine: turning raw voice-to-text into polished prose. That's the Pro tier, because rewriting is a cloud compute call. It's fine to sell that as paid. Cloud-based LLM-driven rewriting has real infrastructure costs. But dictation itself? That's local. That's yours to use as much as you need.
Marcus still uses Cursor because tab-complete cuts rewrites. He still documents his architecture in Notion. He still refuses cloud transcription for code. Now his voice tool doesn't interrupt his thinking anymore. The tool got out of the way. That's the point.
When you're building in a workflow powered by models, the bottleneck is no longer how fast you can type. It's how clearly you can express intent. A tool that meters your dictation by the word is a tool that's optimized for a previous decade. Local-first transcription with no caps solves for the workflow that exists right now.