Published accuracy numbers do not transfer to your voice, your vocabulary, or your meeting room. Here is a seven-day test that measures what actually matters.

Every note taker advertises an accuracy number, and almost none of those numbers apply to you. They are measured on clean audio, standard accents, and a vocabulary that is probably not yours. A tool that scores well on a benchmark can still struggle with your client names, your internal acronyms, and the way you speak when you are tired. The only reliable number is the one you measure yourself.
The good news is that you do not need a lab. You need seven days, a few real recordings, and a stopwatch. The test below is designed to answer a practical question: how much cleanup does this note taker leave me? That question matters more than word error rate, because cleanup is the work you actually do.
Run the test on the tool you are considering, and run it on your own material. Do not use demo scripts. The point is to find the failures that would annoy you in a normal week, before you build a habit around a tool that cannot handle them.
Define the metrics before you start. Word errors count misspelled or wrong words. Structure errors count sentences that need to be rebuilt. Cleanup minutes count the time from reading the raw text to accepting it as usable. If you only track one number, track cleanup minutes, because that is the cost you pay every time you use the tool.
Write down the hypothesis before testing. For example: this note taker will save at least two minutes per note on my vocabulary in a quiet room. A written hypothesis makes the result interpretable. Without it, every outcome becomes a matter of impression, and impressions drift toward whatever you hoped to see.
Start by creating a baseline sample. Record three pieces of real work: a short meeting recap, a paragraph of a document you are writing, and a message you would normally type. Keep each recording between one and three minutes. Longer samples add tedium without adding much signal.
For each sample, save both the audio and the text the note taker produced. Mark three counts: the number of words that are wrong, the number of sentences that need restructuring, and the number of minutes you spend making the text usable. These three numbers are your baseline. Do not average them yet; the categories fail differently.
Use the same microphone and the same room for all three samples. If you normally work in different environments, that is a variable you should test later, not on day one. The baseline should measure the tool, not the room.
Keep the results in a simple table. One row per sample, with columns for errors, structure fixes, cleanup minutes, and a short note about the condition. A table makes patterns visible. You will often find that one condition, such as a noisy room or a fast speaker, produces most of the total cost.
Use your own week as the sample source. Pick notes you would have written anyway, not specially prepared scripts. Real material contains the interruptions, names, and half-finished sentences that determine cleanup time. A test built from clean samples flatters every tool equally.
On days three and four, test the vocabulary that breaks transcription. Make a list of twenty terms you use weekly: client names, product names, internal acronyms, ticket prefixes, and numbers with units. Read the list into the note taker at a normal pace and compare the output.
Count two things: how many terms were wrong, and whether the error was consistent. A tool that misreads the same term every time can be fixed with a personal dictionary. A tool that misreads different terms unpredictably is harder to trust, because you cannot correct it once and move on.
Numbers deserve their own check. Say dates, prices, percentages, and version numbers in the way you normally say them. Transcription errors on numbers are more expensive than errors on ordinary words, because they change the meaning of a note without looking obviously wrong.
Score name errors by term, not by occurrence. If the tool misreads the same client name three times, that is one term with three misses. The distinction matters because a dictionary fixes a term, not an occurrence. A tool with a consistent error pattern is easier to live with than one that fails differently each time.
Count how many terms need a second attempt. A tool that gets a term wrong once and right the next time is easier to live with than one that fails consistently in both directions. Consistency is what a dictionary can fix. Randomness is what forces you to proofread every note, which is the hidden cost the test is designed to expose.
On day five, test the conditions you actually work in. Record the same short passage three times: once in a quiet room, once with background noise such as a fan or a nearby conversation, and once at a faster speaking pace. If you can, add a sample where two people speak close together.
Compare the three outputs for the same failure pattern. Some tools handle noise well but collapse when you speak quickly. Others handle speed but misread consonants in a noisy room. You are looking for the condition that produces the most cleanup, because that is the condition you will meet on a bad day.
Do not test every possible environment. Three conditions are enough to reveal whether the tool is robust or only works in a quiet demo room. If the noisy sample is unusable, that is useful information, not a reason to run more tests.
The noise test should use a realistic distraction. A fan, a nearby conversation, or a television in the next room is more representative than white noise from a speaker. The point is to see whether the tool holds the sentence together when the signal is imperfect, not to simulate a laboratory condition.
Record the condition next to the result. A noisy sample that takes four minutes to clean is useful data only if you know it was noisy. Over time, the log shows which context is expensive, and that tells you where to change the workflow rather than blame the product.
On days six and seven, stop counting word errors and start counting cleanup time. Take each sample, start a timer, and edit the text until you would be comfortable sending it or filing it. Stop when it is usable, not when it is perfect.
Record the minutes for each sample, then compare them with your baseline. The absolute number matters less than the difference. A tool that saves two minutes per note across ten notes a week saves twenty minutes, which is a real change. A tool that saves thirty seconds is a novelty.
This is also where you should notice a subtler cost: the attention required. Some tools produce text that is technically correct but needs constant re-reading to trust. That mental load does not show up in a timer, but it shows up in whether you keep using the tool. Write down your honest impression next to the minutes.
Use a timer and stop at usable. Usable means you would send it to a colleague or file it in the project. Do not polish. The temptation to keep editing inflates the number and makes the comparison unfair across tools, because every tool looks expensive when the standard is perfection.
Keep the raw output and the cleaned note side by side for at least one sample. The difference between them is the clearest picture of what the tool did and did not do. You will often find that the tool handled punctuation and sentence boundaries well, while the remaining work was vocabulary and structure. That split tells you which fix to pursue: a dictionary, a shorter capture, or a different tool.
A good result is not zero errors. It is a pattern you can manage. Terms that are consistently wrong can be added to a dictionary. Numbers that need a second look can be read back during capture. The text should be close enough that cleanup is a quick pass, not a rewrite.
If the baseline is poor and the vocabulary test is poor, the tool is probably not a fit, regardless of its benchmark score. If the baseline is good but the names are consistently wrong, a personal dictionary may close the gap. Voiskey includes one, which turns recurring name errors into a one-time setup task.
Repeat the test after a month of use. Accuracy can change with updates, and your vocabulary changes with new projects. A ten-minute retest is enough to confirm the tool still fits. The goal is not to find a perfect note taker. It is to know exactly where yours fails, so the failure is predictable and cheap to fix.
Retest after a month and after any major update. Vocabulary changes with new clients and projects, and tools change their models. A short retest on two samples is enough to confirm the fit. The goal is a maintained number, not a one-time verdict.
Share the result with the vendor if a consistent term fails. Many tools accept dictionary additions or vocabulary lists, and a specific report is easier to act on than a general complaint. Keep the sample file so you can retest after the fix and confirm the improvement.
To run the test on Voiskey, open the official website and click the download button in the top right corner to get the Mac or Windows app. The app is free to start, and the personal dictionary makes it easy to close the naming gaps the test reveals.
Voiskey is built to catch fillers, stumbles, changes of mind, and grammar slips, which is the kind of speech a real note contains.
Yes. Voiskey learns your names, jargon, and spelling preferences, so recurring terms can be written the way you use them.
Voiskey produces native-sounding text in 100+ languages, and you can speak your own language when you need a translation.
Use a real note in the app you normally write in and count the cleanup time. Voiskey adjusts formatting, tone, and word choice to your intent, and the timer shows whether that brings the text close to ready to send on your material.

Download Voiskey for Mac or Windows and run the seven-day test on real notes and real names. Free to start.
Use it across macOS, Windows, iOS, and Android