Speech-to-text apps: which to pick

A speech-to-text app, in one line: you talk, it types

Here is the whole mechanic, stripped of marketing. You press a key. You speak. The app records your voice, runs it through a recognition model, and the words appear as text. With our app, Whisper by Remskill, the default is Ctrl+Space on Windows and a command+option hold on macOS: press, talk, release, and the transcript pastes at the cursor in whatever app you were already in.

That last part is the part most people miss when they're comparing apps. A speech-to-text app is only as useful as the place it can put the text. Apple Dictation types into Mac and iOS apps. Windows voice typing opens with Windows logo key + H and works in any text field on a PC. Google Docs Voice Typing only works inside Google Docs, in Chrome. They all transcribe; they don't all go everywhere. So before you compare accuracy scores, ask where you need the text to show up.

How a speech-to-text app works (ASR plus the cleanup pass)

Underneath, every speech-to-text app does two separate things, and conflating them is where reviews go wrong.

The first is transcription: automatic speech recognition turns the audio waveform into a string of words. This is the hard part, and it's where model quality shows. NVIDIA's Parakeet model, which we ship as one engine option, posts a 6.34% average word error rate on the Hugging Face Open ASR Leaderboard. Dragon Professional advertises "up to 99% recognition accuracy, right from first use." Numbers like that are real, but they describe clean audio in a quiet room. Background noise and a cheap laptop mic move them more than any model swap does.

The second is the cleanup pass: turning "send her the file um the second version not the first one" into a sentence with a capital letter and a period. Better apps run an extra AI step here to fix punctuation, casing, and the occasional "um." That cleanup is also where dictated commands get interpreted, where filler gets cut, and where a wall of run-on words becomes something you'd send. Transcription gets you the words. The cleanup pass gets you text you don't have to retype.

What separates a good app from a typing simulator

Most speech-to-text apps clear the basic bar: they get the words mostly right. What separates a tool you keep from one you uninstall in a week comes down to a few things that rarely make the feature grid.

Latency, first. If there's a long pause between you finishing a sentence and the text appearing, you lose your train of thought and start re-reading instead of talking. Accuracy on your voice, second — not the demo voice, yours, with your accent, in your room. Language coverage, third, if you're not working in English. And reach: can it type into the one app you live in, or only into a browser tab.

Then there's the boring one nobody markets: where your audio goes. A built-in OS tool and a local app keep the recording on your machine. A cloud app uploads it. For a grocery list, who cares. For a legal draft or a patient note, it's the whole decision. We'll get to that.

What Whisper by Remskill looks like

Since I keep mentioning our app, here's what it is so the comparison is fair. Whisper by Remskill is a system-wide desktop app — not a browser extension, not a Word add-in, not a meeting bot. You hold the hotkey, speak, release, and the transcript lands at your cursor in any app or text field.

For transcription you pick one of three paths. Local Whisper runs eight model sizes from Base (~140 MB) up to multilingual Large v3 (~3 GB), supports 99 languages on the multilingual variants, and can translate to English. Local Parakeet (~600 MB) is 5-10x faster than Whisper on CPU and covers English plus 24 European languages. Cloud mode is bring-your-own-key OpenAI for people who want the latest models and live web answers. All local transcription runs in pure Rust, fully on-device, and works offline. The app doesn't pick for you. It shows the three paths and you choose based on speed, languages, or accuracy. For the full list of how dictation tools compare, our guide to transcription software walks through the same trade-offs across more tools.

Deciding between the two local engines is the most common question we get, and our Whisper vs Parakeet explainer breaks down speed, language coverage, and accuracy side by side.

Cloud vs. local: where your voice goes

Here's the one I have a strong opinion about. Cloud-only speech-to-text is a privacy problem most people don't think about until it's too late. Your boss's salary spreadsheet, the email to your kid's school, the draft of a contract — none of that should sit in a vendor's logs because you wanted to type with your voice.

I once watched a team at a company I worked with rack up a five-figure cloud-AI bill in a single quarter, because a "smart retry" loop kept re-transcribing the same standup recordings four times over. The CFO opened the dashboard at the quarterly review and the room went quiet. The cost was the headline. The quieter problem was that every recorded meeting had been shipped to a third-party API to find that out.

Local-first solves both. Our local pipeline never sends your audio anywhere; Cloud mode is the opt-in exception, and it's your own OpenAI key, not ours. Wispr Flow, the closest direct competitor, is cloud-only. Audio leaves the device by design. Neither approach is wrong. But if your work touches anything you'd hesitate to email, run it on your own machine.

Raw words in, clean text out: the AI cleanup step

The cleanup pass deserves its own mention because it's the difference between transcription and usable text. Recognition gives you the words in order. It doesn't always give you a sentence.

In local mode, our app runs the optional cleanup through Ollama on your machine — punctuation, casing, and trimming the verbal filler, all without an internet round-trip. In Cloud mode the same job goes to OpenAI's models via your key. The point isn't which model does it. The point is that a speech-to-text app that skips this step hands you a paragraph with no commas and a stray "you know" every third line, and you spend the time you saved on cleanup. The good ones do that work before the text reaches your cursor. The basic built-in tools mostly don't, which is fine for a quick note and frustrating for anything longer.

The other speech-to-text apps worth knowing

No single app wins every use. Here's the honest field, and who each one is for.

Late in the 1990s a relative of mine ran Dragon NaturallySpeaking on a Windows 98 desktop with 64 MB of RAM. The setup made you read a word list for 45 minutes to "train" it, then dictation worked at maybe 70% accuracy with a four-second delay per sentence. Dictating one paragraph of a holiday letter took fifteen minutes. The headset got thrown across the room; it survived, the experiment did not. Twenty-five years later a seven-year-old can dictate a full email in ninety seconds without training anything. If a speech-to-text app still asks you to "train" it before it works, it's living in 1999.

When to skip a dedicated app entirely

Here's where I send you elsewhere (yes, the guy selling the app). If you only dictate short notes — a text, a calendar entry, a quick reply — and you're on a Mac or iPhone, just use Apple Dictation. It's free, it's built in, and for under a hundred words it does the job; no hard timeout on modern macOS. On a PC, Windows logo key + H does the same job for free.

And if you land on the install-something side, picking a voice typing app for each device is its own short guide.

And if all you ever do is record meetings and want a searchable transcript afterward, a dedicated meeting transcriber like Otter is the right category. It connects to Zoom and Teams and is built for that exact job. A system-wide dictation app starts earning its keep around the point where you're writing longer text, across many apps, and you care about where your audio goes.

Free vs. paid: what $0 gets you

The free options are real, not crippled trials. Apple Dictation, Windows voice typing, and Google Docs Voice Typing cost nothing and handle everyday dictation. With our app, the entire local pipeline — Whisper, Parakeet, on-device AI cleanup, history, presets — is free for any signed-in user, with no card required to start.

Paid tools earn money on top of that: cloud features, more languages, team management. So the real question isn't free versus paid. It's whether the free tier you're handed reaches every app you work in and keeps your audio on your machine. If it does, stop there. When my wife asked why I'd give away the whole local pipeline, I gave her a five-minute answer; the short version is that I'd rather you start free and only pay if you hit a wall. Pricing for the cloud tier lives on our pricing page.

My younger daughter's verdict on all of this was simpler than any review. She pressed the key once, said a sentence about a tooth she'd lost, watched it appear, and went back to drawing without a word. No accuracy chart, no feature grid, just text where she wanted it. That's the only test that ever mattered: did the words show up where you needed them, and did you have to think about it. If the answer's yes and no, you picked the right app.