To start, choose the sound-wave icon. The microphone icon is dictation, which turns speech into text input; voice mode also speaks its answers. Claude's current beta supports web, Desktop, iOS and Android, with availability and connected tools subject to account and organization conditions.
Make one short request before a long discussion
Start in a quiet place and ask something easy to evaluate. Try this practice prompt:
Help me think through the order of three tasks for a practice event.
I must check the room before setting out chairs. The printed signs
can be prepared independently. Ask one question at a time about
missing constraints. Do not book anything or use connected tools.
You should be able to finish a sentence, hear a reply, and tell whether the reply used the stated dependency. A proposed order that puts chairs before the room check needs correction, regardless of how natural the voice sounds.
Choose the speaking control that fits the room
In hands-free mode, Claude listens continuously and responds to pauses. You can interrupt a reply by speaking again. If nearby conversation triggers false interruptions, use push-to-talk rather than repeatedly restarting your explanation. Speaking controls
Consider three different observations:
| What you notice | First change to try | What to observe afterward |
|---|---|---|
| Claude answers during your thinking pauses | Hold push-to-talk through one complete thought | Whether the answer starts after you release the button |
| Another person's speech becomes input | Move away from that speaker and use push-to-talk | Whether the transcript contains only your intended words |
| The reply is audible but breaks up | Check the connection and other resource-heavy apps | Whether the same short request produces continuous audio |
Change one thing, then repeat the same harmless request. That makes it easier to distinguish a listening problem from a playback or connection problem. Do not use repeated readings of confidential material as your microphone test.
Type details that speech keeps getting wrong
You can switch between text and voice in the same conversation. Use text for a URL, reference number, formula, or spelling that must survive exactly. Text and voice continuity
For the event example, imagine that the room label is C-14B. If the transcript repeatedly shows C-40B, type the correct label and say:
Use the room label I just typed: C-14B. The spoken transcription
was wrong. Keep the room-check dependency attached to C-14B.
Then inspect the next written result. Repeating the label aloud is less useful if the same error keeps returning. A fluent response does not establish that the reference was corrected everywhere.
Check the language setting separately
For a conversation in another language, set Settings > General > Voice > Language. Changing the app's display language does not change the voice language. Anthropic describes non-English voice support as beta. Language settings
Try a short sentence in the selected language before dictating a complicated request. Include an ordinary sentence and one important term. If the ordinary sentence is understood but the term is wrong, type that term; if the whole sentence is misrecognized, recheck the selected voice language and microphone conditions.
End with the checked result you need
Ask for a short written task order and inspect it: room check first, chairs afterward, signs independent, and C-14B preserved. Stop voice mode with the displayed Stop control when you are finished.
Voice transcripts remain in chat history. Review the relevant account controls before choosing this route for sensitive work. For a spoken prompt that you want to inspect before sending, Claude dictation on Mac covers that separate input method. Keep the final task order where the event team normally maintains its plan.