sayhushTest a dictation app with several short samples from your real work, using the same microphone and a clear reference text. Track meaning-changing errors, editing effort and whether text reaches the right field. This is a test method, not a published benchmark.
Decide what a useful result means
Write down the jobs you expect the app to do: a daily note, an email draft or an explanation for a colleague. Define the review standard for each. A personal reminder may tolerate rough punctuation; a delivery instruction needs the right quantity and date. This makes the test more informative than comparing two unrelated demo transcripts or a vendor’s headline accuracy claim.
On a small screen, scroll the table sideways to see all columns.
| Record | Example value to replace |
|---|---|
| App / version / mode | Your installed version and chosen mode |
| Input / environment | Microphone, room and language |
| Sample / reference | Your original text and what you actually said |
| Result | Meaning-changing errors and correction time |
| Insertion | Destination and whether the text arrived correctly |
Create a small original sample set
Prepare four short samples: ordinary prose, names and numbers, vocabulary from your work, and a realistic language sample. Use fictional or public details so the test does not expose private information. For example: “Ask Noor to review 15 draft pages by Tuesday.” That sentence is a suggested test input, not evidence of how any product transcribes it.
Keep the conditions comparable
Use the same room, microphone, speaking distance and destination editor. Note the app version, input language and any processing mode. If an app rewrites text while another preserves a transcript, label that difference before scoring. Read naturally and record departures from the reference. Otherwise a word you did not actually say could be incorrectly counted as a recognition failure.
Record errors that affect the task
Mark wrong words, missing words and additions, then identify which ones change the meaning. Keep cosmetic preferences separate from factual mistakes. A transcript that needs one corrected amount can be more consequential than one with several harmless punctuation differences. Record how long it takes to make the draft usable, and do not turn a small informal sample into a scientific accuracy ranking.
Test insertion and interruption separately
After checking transcription, test the rest of the workflow: focus the target field, start, pause, resume and finish. Note whether text lands where expected and whether you can recover from a cancelled session. A correct transcript that cannot reach your editor is a workflow issue. Keep that separate from recognition quality so you know what needs fixing.
Repeat before choosing a winner
Run more than one sample on more than one occasion. Compare the pattern of errors and the finished editing effort, not the best single result. Publish results only with the actual samples, conditions and limitations. Sayhush’s free allowance can support a small trial, but this guide does not claim that it outperforms another product or every built-in option.