1. Check that your account can transcribe at all
The most confusing failure looks like a working feature: the mic button lights up, Cursor records, and no text ever appears. In a May 2026 thread in Cursor’s forum, a Cursor staff member explained that voice input and transcription are only available on paid plans, Pro and above. On a Free account the audio is sent for transcription and rejected on the backend, and the app did not say so clearly; the same reply called that bad UX and said it had been reported to the team.
So before reinstalling anything, check which plan the account signed in to the app is on. If you have more than one Cursor account, make sure the paid one is the one Cursor is using.
2. Grant microphone access to Cursor itself
Cursor needs operating-system permission like any desktop app. On a Mac, enable Cursor under System Settings → Privacy & Security → Microphone. On Windows, turn on microphone access and access for desktop apps under Settings → Privacy & security → Microphone, and check which device is the default recording device in Sound settings; Cursor staff ask for exactly that when diagnosing Windows reports.
A microphone that works in Chrome proves the hardware, not Cursor’s permission. Watch the waveform around the button while you speak. Cursor staff use it to separate two problems: a flat meter means audio is not reaching Cursor, while a moving meter with no text means audio arrives but transcription fails.
- Flat meter: permission, input device, or a client bug.
- Moving meter, no text: plan, network, or Cursor’s transcription backend.
- Text that appears, then rewrites itself: see section 5.
3. Know which of Cursor’s two voice paths you are using
Cursor 2.0, released October 29, 2025, introduced Voice Mode: built-in speech-to-text for controlling Agent, plus custom submit keywords you define in settings to start the agent. Cursor’s keyboard reference lists Cmd+Shift+Space as Toggle Voice Mode, and its prompting docs describe clicking the microphone icon in the chat input, speaking naturally with file and function names included, and reviewing the transcription before sending.
Cursor 3.1, released April 13, 2026, upgraded voice input in the Agents Window: press and hold Ctrl+M to speak, with a waveform, a timer, and buttons to cancel or confirm, and the full clip is transcribed with batch speech-to-text. Cursor staff later described the Agents Window as a separate code path that inserts the text as one block after you stop.
When one path fails, try the other. During editor-layout problems in May 2026, staff recommended the Agents Window mic because transcription worked there on the same build. You can see which layout you use under Cursor Settings > General > Layout.
4. The shortcut does nothing, or the mic button is gone
If Cmd+Shift+Space or Ctrl+M does nothing, open Keyboard Shortcuts (Cmd+R then Cmd+S, or the command palette entry Keyboard Shortcuts), search for “voice”, and check that the binding exists and is not taken by another command. Cursor says all of its keybindings, including Cursor-specific features, can be remapped there.
A missing microphone button is often a state, not a fault. Cursor staff confirmed in May 2026 that once the input holds text or a file reference, the mic is replaced by the send arrow, so you cannot start another recording until you send or clear the input. They described it as a known UX issue under tracking. If you rely on the submit keywords Cursor 2.0 introduced and a keyword does not send the prompt, press Enter yourself rather than recording again.
5. Audio is captured but text is garbled, cut off, or never arrives
Cursor transcribes on its servers. In May 2026 a Cursor staff member said a voice transcription failure had been “fixed on the backend” with no client update needed, and later referred to it as a Baseten outage fixed server-side. Cursor’s voice documentation does not say where audio is processed or how long it is kept, so treat voice like any other prompt you send to Cursor’s service.
For text that appears, then rewrites itself or drops chunks during longer dictations in the editor, staff attributed the behavior to progressive speech-to-text and pointed to the Agents Window, where text arrives as one block after you stop. For anything else, collect evidence before you file a report.
- Run Cursor Settings > Network > Run Diagnostics.
- Open Help > Toggle Developer Tools > Console, record a few words, stop, and copy the output.
- Note your Cursor version, operating system, layout and microphone type (built-in, USB or Bluetooth).
- Post it in Cursor’s Bug Reports forum, with a request ID if you have one.
6. Dictate into Cursor without its voice backend
Cursor’s own voice input is the shortest path when it works and your plan includes it. When it does not, because of a Free plan, a backend outage, or a prompt you want spoken somewhere other than the chat input, a system-wide dictation tool types into whichever Cursor field has focus. Cursor staff themselves suggested macOS system dictation as a stopgap that types straight into the chat input.
IraVoice 1.2.2 is a dictation app for macOS 13+ on Apple silicon, Windows 10 or 11 (64-bit), and Linux (beta). Recognition runs on your computer with NVIDIA Parakeet, so Cursor’s plan and servers do not affect it, and on IraVoice Local no audio or text leaves the machine. Hold Right Command on a Mac or Right Ctrl on Windows with the agent input, an editor tab, the integrated terminal or the commit box focused, speak, and release; nothing is sent until you press Enter.
Cursor is on IraVoice’s list of recognized apps on both Mac and Windows, so it types names from the Git project Cursor has open the way the code spells them, such as useEffect for “use effect”. On Linux, IraVoice is a beta and dictates into Cursor without code names. IraVoice Local is $79 once for one computer, after a free trial.