The problem with voice-to-text for coding

I dictate a lot. Thousands of prompts a month into Claude, Cursor, Copilot, whatever I’m using that day. Every voice tool I tried had the same problems.

They’re all cloud-only. My audio gets shipped to some server, transcribed, shipped back. The latency alone kills any flow state. Most of them want a monthly subscription too, sometimes per-minute pricing, just to use my own voice. And none of them understand code. I say “create a POST endpoint for user auth” and get back create a post endpoint for user off. Useless.

I was spending more time fixing transcription errors than I saved by speaking in the first place.

What I tried before building anything

I didn’t start by writing code. I started by paying for things.

Wispr Flow has the right interaction model, but it’s cloud transcription on a $12/month subscription, and the Windows app trails the Mac one. SuperWhisper runs Whisper locally, which is right, but it’s Mac-first and the Windows version is limited. Talon is genuinely impressive as a full voice-control system, but I didn’t want to learn a command grammar, I wanted to talk.

And I tried the free route: raw faster-whisper in a Python script with a global hotkey duct-taped on. That’s when it clicked that the model was never the problem. The missing piece was everything around it: the push-to-talk loop, the auto-paste, and something to turn rambling into a usable prompt.

What I actually wanted

Something simple: hold a key, speak, release. Get sub-second transcription on my GPU, locally. Have the output land at my cursor, already formatted for the tool I’m using. No subscription, no cloud, no word limits.

That’s it. That’s the whole product.

How it works

Invoke runs faster-whisper on your GPU. Hold your push-to-talk key, it records. Release, it transcribes. Under a second on most modern GPUs.

Raw transcription isn’t enough for coding prompts though. “Add a post endpoint that takes a keyword and URL generates an article saves to DB skip auth for now” is what Whisper gives you. Accurate, but not what you’d actually type into Claude or Cursor.

So there’s an optional AI reformatter. It takes the raw transcript, reads your project context, and rewrites it as a clean prompt. The kind you’d write if you sat down and thought about it for thirty seconds instead of rambling into your mic.

Output gets auto-pasted at your cursor. Speak, release, done.

The feature I didn’t plan: screenshot-to-path

The second thing Invoke does wasn’t in the original idea at all. I kept hitting the same wall dictating into Claude Code: terminals can’t paste images, so every screenshot meant save, find the file, type the path. I added a hotkey that grabs the clipboard image, saves it as a PNG, and pastes the file path into the terminal. That one feature now drives half the traffic to this site, because it turns out everyone fights that wall.

The lesson I took: the product isn’t “transcription.” It’s getting things into the prompt fast, whether that thing is your voice or a screenshot.

> Sounds interesting?

Invoke is $49 once. Free 7-day trial, all features, no credit card.

Try it free →

Why local matters

Transcription happens entirely on your GPU. No internet, no audio leaving your machine.

The AI reformatter does call an LLM API (you pick which one), but it’s optional and only sends text, never audio. If you want the whole pipeline local, point it at Ollama or LM Studio and you’re done.

The pricing decision

$49 once, not a subscription, and that was a deliberate call, not a growth hack. Dictation is an input method. You don’t rent your keyboard. The cloud tools charge monthly because they pay for GPU compute on every utterance you make; Invoke runs on your GPU, so there’s no meter running and nothing to subscribe to. I wrote up the full no-subscription math separately, but the short version: a $12/month tool costs more than Invoke by month five, and it never stops.

What’s next

Invoke started Windows-only. The macOS build shipped in August 2026: Apple Silicon (M1 or newer), transcription on the M-series GPU through MLX, signed and notarized by Apple. Same $49, same 7-day trial, either platform.

If you spend your day talking to AI dev tools, try it free. 7-day trial, all features, no credit card. Worst case you go back to typing.