
Phone dictation is most useful when it saves effort without giving up control of the message. Speak a short passage, inspect the result, correct the important details, and only then send it. Automatic punctuation and fluent wording are not proof that the system heard you correctly.
Dictation, captions, and summaries do different jobs
Dictation puts recognized words into an editable field: a message, note, or document. Live captions or transcription help someone follow speech as it happens or review a recording. A summary selects and rewrites information. It is a derived account, not a word-for-word record.
This distinction changes what to check. For a dictated message, you want the correct recipient and intended wording. For a transcript, you may also need speaker identification and timestamps. For a summary, compare the claims against the underlying conversation, especially decisions, unresolved questions, and who agreed to do what.
Google's Live Transcribe documentation describes accessibility features and offline-language support on eligible devices. That is a different workflow from inserting text with the keyboard microphone. Do not assume that a keyboard's voice-input setting also controls an app's captions, recording, or transcription storage.
Start in a draft you can safely edit
On iPhone, go to Settings → General → Keyboard and enable Dictation, then use the microphone when a text field and keyboard are open. Apple's Dictation guide explains how voice input and touch editing work together. Available commands and languages vary, so consult the guide for your system version.
In Gboard, open a text field and use the keyboard microphone. Basic voice typing and the advanced Pixel features are not the same feature set. Google's advanced voice-typing guide specifies compatible devices, languages, and settings; the advanced workflow requires Pixel 6 or later. Do not assume every Android keyboard has the same commands.
- Choose the language and destination. Open a private test note rather than a live conversation. Check that the keyboard language matches your speech.
- Speak a complete short thought. Use an ordinary clear voice, with pauses between sentences. Try automatic punctuation first, then use supported spoken punctuation commands if needed.
- Stop input before reviewing. A microphone that remains active can add background speech or interpret further words as commands.
- Inspect meaning before style. Check people, amounts, dates, locations, and words such as “not.” Then fix punctuation and formatting.
- Confirm the recipient and send deliberately. A correctly transcribed message sent to the wrong person is still a failure.
Keep frequently needed names and specialist terms available to paste or type. Repeating an unrecognized name louder is often less efficient than correcting it once. If an app supports vocabulary personalization, review its documentation and privacy controls before enabling it.
Try this original twenty-word check
The text and faulty result below are invented examples, not measurements of any product. Read the reference into a draft, then compare your actual result before making corrections. For this exercise, count space-separated words and ignore capitalization and punctuation.
The reference has 20 words. The invented result replaces “four” with “five” and omits “green”: one substitution and one deletion. Using (substitutions + deletions + insertions) ÷ reference words, the word error rate is 2 ÷ 20 = 10 percent. There are no insertions in this example.
That percentage hides the practical consequence: the time is wrong and one folder is no longer identified by color. A passage with very few errors can still be unusable if the errors change its instructions. This short exercise is a personal setup check, not a statistically meaningful accuracy benchmark or a fair ranking of speech systems.
Download the reference and scoring notes (plain text) and dictation check worksheet (CSV). Repeat with the same device, language, and passage in a quiet room, then with your usual microphone arrangement. Record uncorrected output separately from the version you edited. Measure correction effort as well as whether the microphone recognized speech.
After that controlled check, try a short original sentence containing the names and vocabulary you actually use. Avoid personal or confidential details while learning how the feature processes and stores input. Success with this English passage does not establish accuracy in another language or with a different speaker.
On-device recognition does not make every later step local
Apple says Dictation is processed on the device for many languages, but notes that text dictated into a search box may be sent to the search provider. A note app may also sync the resulting text. Verify both the recognition feature and the destination app; these are separate data paths.
Google says advanced Gboard voice typing normally keeps the dictated text on the device, but identifies exceptions for “Fix it” and detailed editing features. Its documentation says detailed edits send the command transcript and input-field text to Google for processing. A promise about the basic dictation step should not be extended to every optional AI tool.
If offline use matters, download any required language resources in advance. Then test with a non-sensitive note while both mobile data and Wi-Fi are off, and restore your normal connections afterward. Passing this test shows that the tested action can work without a live connection; it does not prove that the app never syncs content later or that every supported language works offline.
Before transcribing another person, agree on the purpose, who will receive the text, and how long it will be kept. Check any applicable recording rules. Automatic speaker labels and summaries can be wrong, so retain a way to verify an important decision through an authorized record or direct confirmation. See Cell Phone Privacy for the treatment of copies, accounts, and linked devices.
Fix the symptom you actually have
Scroll sideways on a small screen to read the troubleshooting table.
| Symptom | First check | Useful next experiment |
|---|---|---|
| Nothing appears | Microphone permission, active text field, and selected input language. | Try a short phrase in another draft field. |
| Words are consistently wrong | Language choice, microphone position, and competing sound. | Repeat the same passage in a quieter setting. |
| Commands become literal words | Whether that command is supported in the current feature and language. | Use touch editing, then test one documented command. |
| Works online but fails offline | Feature eligibility and downloaded language support. | Use the supported offline configuration, or plan for connectivity. |
| Text looks fluent but changes meaning | Whether an AI rewrite or correction was applied. | Compare with the original draft and restore the intended wording. |
A transcription workflow succeeds when you can reliably check and correct it. For a quick social message that may take seconds. For a contract term, address, or consequential instruction, confirm the detail directly instead of relying on a smooth-looking sentence.