- You dictate an email, and the client's name comes out spelled three different ways. The reflex is to look for better software. That is almost always the wrong first move.
- A dictation tool is only as good as the audio it receives and the words it already knows. Both are fixed in ten minutes, without paying anything.
- This article explains why the microphone and the vocabulary matter more than the model, then gives you a test to check it on your own PC.
You dictate an email to a client. The first two sentences come out clean, then the client’s name shows up with three different spellings, and your acronym turns into a common word. You spend longer correcting than you would have spent typing. The reflex is to look for better software. I think that is the wrong first move.
A dictation tool is only as good as the audio it receives and the words it already knows. Both of those are fixed without spending anything.
The model cannot recover what the microphone lost
A speech recognition model transcribes the signal it receives. It guesses words from the sound, then picks the most likely one for the language. If the sound is muffled, echoing, covered by a fan, or heavily compressed, information is missing, and the model cannot invent it. It guesses more often, and its guesses sound plausible. They are harder to spot when you proofread.
A Bluetooth headset shows the problem well. When its microphone is in use, the connection switches to a low-bitrate voice mode, and the sound quality drops sharply compared with music. The same tool, fed by a USB microphone placed near your mouth, produces a more faithful text. This is not about brand. It is about signal.
My view is clear: before you pay for a more expensive tool, plug in a wired microphone, even a modest one. It is the purchase that pays back fastest, and often the only one that really changes the result.
Vocabulary fixes a pattern, not a one-off error
A wrong spelling of a client’s name never happens just once. It comes back at every dictation, because the model does not know that name. Correcting the same word by hand ten times is the worst return you can get from dictation. A tool that accepts a personal vocabulary fixes that pattern once.
That vocabulary should hold what the model cannot guess: proper names, your industry’s acronyms, product names, the terms a general tool keeps misspelling. Common words do not need to be added. A short list, reviewed once a month, is worth more than a two-hundred-word list nobody maintains.
If your tool offers no vocabulary, a replacement rule does the same job: the misspelled word becomes the right form every time.
Dictate in segments, not in one go
There are two ways to dictate. Continuous dictation shows words while you speak, the way Aqua Voice does. It is reassuring, but it asks you to watch the screen while talking, and live corrections break your train of thought. Segment dictation lets you finish an idea before you look at the result.
For an email or a note, I dictate in chunks of two or three sentences. That is short enough to keep corrections simple, and long enough to keep the thread. Dictating a whole paragraph in one go usually produces wordy text that you barely proofread, because by the end you no longer remember what you meant at the start.
Clean up with AI afterwards, but keep it away from the numbers
Some tools add an AI layer that rephrases, punctuates, and smooths the transcript. That is useful for a first draft. It is risky for three kinds of content: figures, proper names, and commitments. An AI that “improves” a sentence can change an amount or a date without telling you.
My rule: transcribe first, clean up second, and check the figures, names, and dates yourself. Your own proofreading remains the only reliable control.
A ten-minute test
To find out which factor costs you the most errors, run this test once, on your own PC.
- Read aloud a hundred-word text you know, with two proper names and one acronym. Dictate it with your current microphone, then count the errors: wrong words, names spelled differently, missing punctuation. Write down the total.
- Plug in a USB microphone or a wired headset, and repeat the exact same reading, at the same distance from your mouth. Compare the totals.
- Add the two proper names and the acronym to the tool’s vocabulary, then read the text a third time.
- Keep the setup that makes the fewest errors. Change tools only if, after these three tries, the gap still bothers you.
For a free, local option, Handy is open source and runs on Windows, macOS, and Linux. You can test it with this same protocol without buying anything.
What this changes for you
This week, run the test once, then keep the setting that wins. If you want the full method, covering microphone choice, the activation key step by step, the comparison of tools, and the short proofreading pass, the Voice dictation lesson in the Windows module details it. For the Windows settings that make the rest of your day easier, the free guide to 10 Windows settings is the right starting point.
Sources
Last verified: November 1, 2026. Handy’s presence and features were rechecked on its official site; the rest of the article rests on no figure.
See the source for this article
- Handy, official site, handy.computer, open source speech-to-text application