gemini-cli - 💡(How to fix) Fix Voice Mode (gemini-live) fails with 'Connection Closed. Code: 1000' despite correct audio format

What happened?

When activating /voice mode, the Gemini CLI attempts to connect to 'gemini-3-flash-live-preview' via WebSocket but immediately logs '[Voice] Stop requested' followed by 'Connection Closed. Code: 1000, Reason:'. The voice mode does not function at all. This occurs consistently, even after extensive troubleshooting to ensure the audio input device provides the correct sample rate and my API key is linked to a project with a prepaid with enough funds and a postpaid plan using Google AI Pro (5 TB) monthly subscription.

What did you expect to happen?

I expected the /voice mode to successfully connect, transcribe my audio input, and provide a response.

Client information

Run gemini to enter the interactive CLI, then run the /about command.

> /about
│ About Gemini CLI                                                                                                                                                │
│                                                                                                                                                                 │
│ CLI Version                                             0.44.0-nightly.20260515.g928a311fb                                                                      │
│ Git Commit                                              ed9bcf46b                                                                                               │
│ Model                                                   gemini-3-flash-preview                                                                                  │
│ Sandbox                                                 no sandbox                                                                                              │
│ OS                                                      win32                                                                                                   │
│ Auth Method                                             gemini-api-key

Login information

API Key

Anything else we need to know?

Extensive troubleshooting has been performed, suggesting an underlying issue with audio input handling or service initialization, even when the audio format is correct. Steps taken include:

• Verified 'experimental.voiceMode' is true and 'experimental.voice.backend' is 'gemini-live' in settings.json.
• Attempted both 'push-to-talk' and 'toggle' activation modes, both of them open the recording mode without instant crashing. The recording stays ON as long as I press/toggle. The --debug mode F12-window shows this:

  [GeminiLiveTranscription] Connecting to model gemini-3.1-flash-live-preview via raw WebSocket with API Key...
  [Voice] Stop requested
  [Voice] Draining transcription for 4000ms
  [Voice] Grace period ended, disconnecting service
  [GeminiLiveTranscription] Connection Closed. Code: 1000, Reason:

• Configured VB-Audio Virtual Cable (CABLE Input) as the default recording device in Windows Sound Settings, explicitly setting its Default Format to '1 Channel, 16 bit, 16000 Hz (Tape Recorder Quality)'.
• Confirmed the virtual cable's output with FFmpeg: Initial detection showed 44.1kHz stereo, but FFmpeg could successfully resample it to 16kHz mono. However, even with the virtual cable configured, the CLI exhibited the same behavior.
• No explicit audio device errors (e.g., WASAPI errors) were visible in the F12 debug console during 'gemini --debug', only the 'Stop requested' lifecycle message.