Unfamiliar names and internal vocabulary
Speech models are good at ordinary language and much less certain about a new product name, an internal acronym, or a dependency with unusual spelling. The result can sound right while being textually wrong.
Add recurring terms to your vocabulary, say ambiguous names with a little more space around them, and scan proper nouns before sending.
Dense code and punctuation
Natural-language descriptions of code work well. Dictating exact code often does not. Brackets, operators, casing, indentation, and short identifiers create too many opportunities for a plausible but incorrect result.
Use voice to explain the change or compose the agent prompt. Use the keyboard for the exact patch, command, regular expression, or database migration.
Noise, distance, and changing microphones
A laptop across the room, a mechanical keyboard, overlapping speakers, and sudden input-device changes all reduce the quality of the signal. Better language processing cannot reliably restore words that were never captured clearly.
Keep the microphone close, confirm the selected input device, and retry short sections instead of continuing through a bad recording.
The model cannot see every context
A cleanup model can improve a sentence without knowing the private joke, project decision, or unstated tone in your head. It may make an intentionally casual message too formal or flatten wording you wanted to preserve.
Use transcription-only output when exact phrasing matters. For refined output, speak the audience and desired tone aloud when they are not obvious from the destination.
Permissions, focus, and network failures
System-wide dictation depends on microphone permission, accessibility permission, and a stable destination field. It also needs a working connection because Mach uses cloud transcription. If any link in that chain breaks, the correct behavior is a clear failure or cancellation rather than mystery text appearing later.
When a result does not arrive, check the HUD state, connection, input device, and the currently focused app. Mach’s support guide covers the recovery paths.
Review anything consequential
Dictation is an input accelerator, not a substitute for judgment. Read legal, medical, financial, security-sensitive, or irreversible instructions before sending them. The same rule applies to typed text, but faster input makes a deliberate review habit more important.