Voice mode

Voice mode lets you speak your prompts instead of typing them. Your speech is transcribed locally with OpenAI Whisper (via whisper.cpp) and, by default, submitted to the assistant automatically. You can switch to review-before-send if you prefer. It runs entirely on your machine by default, so no audio leaves your computer and no cloud API key is required. The speech model downloads once, on first use.

1. Start a recording with /voice or by holding Space.
2. Speak your prompt while the recording indicator shows your audio level in real time.
3. Stop by pressing Enter, or let it auto-stop after a period of silence.
4. The transcription is submitted to the assistant automatically by default, or placed in the prompt for review when you set voice.autoSubmit false.

Recording, press ENTER when done... 3.5sec ░░░█████████

Partial text streams into the input as you speak, so you can watch the transcription form before you stop.

/voice

Starts a single recording session in the current chat.

Hold the Space key at the prompt for 1.5 seconds to start recording. Release Space to stop and transcribe. Useful for a quick voice input without pressing Enter.

The Whisper model is downloaded on first use, and the download is never silent.

If you trigger voice with Space (push-to-talk) before the model is downloaded, Kiro prompts you to run /voice for the interactive setup.

Local transcription uses whisper-rs (Rust bindings for whisper.cpp) and needs no internet after the initial model download. Two model sizes are available:

Set the model size:

kiro-cli settings set voice.modelSize small

When you work on a cloud desktop with no microphone, voice mode can stream audio to a small server running on your local machine. Voice mode falls back to the remote server automatically when no local microphone is detected.

Manual setup:

# On your local machine (the one with a microphone):
kiro-cli voice-serve

# Open a reverse tunnel when you SSH to the cloud desktop:
ssh -R 19876:localhost:19876 cloud-desktop

# On the cloud desktop, point voice mode at the tunnel before starting Kiro:
export KIRO_VOICE_SERVER_URL=http://127.0.0.1:19876
kiro-cli

You can also force the remote backend for a single session with /voice --model remote (the default is --model local-whisper).

From your local machine, voice-cloud-setup starts the local voice server and opens the reverse tunnel:

kiro-cli voice-cloud-setup <cloud-hostname>

Then, on the cloud desktop, set the server URL in the same shell before starting Kiro and activating voice mode:

export KIRO_VOICE_SERVER_URL=http://127.0.0.1:19876
kiro-cli

The helper verifies SSH connectivity, starts voice-serve locally, and opens the reverse tunnel. It does not currently configure voice.serverUrl on the cloud desktop.

Voice mode uses Silero Voice Activity Detection to tell speech from background noise. It detects when you start and stop talking, filters out keyboard clicks and fan noise, and drives the silence auto-stop, which prevents false transcriptions from ambient sound.

Recent conversation context is also passed to Whisper's initial prompt, which improves accuracy for code identifiers, function names, and project-specific vocabulary discussed earlier in the session.

Voice behavior is controlled through kiro-cli settings. All settings are optional.

Examples:

# Favor speed
kiro-cli settings set voice.modelSize base
kiro-cli settings set voice.silenceTimeout 3

# Favor accuracy
kiro-cli settings set voice.modelSize small
kiro-cli settings set voice.silenceTimeout 8

# Review each transcription before it is sent
kiro-cli settings set voice.autoSubmit false

# Transcribe Spanish
kiro-cli settings set voice.language es

# View your voice settings
kiro-cli settings list | grep voice

• With local Whisper, audio is captured, transcribed, and discarded on your machine. Nothing is sent to any server.
• The remote voice server option streams audio only to the server URL you configure (typically your own laptop over an SSH tunnel).
• Kiro verifies the Whisper model file against a pinned SHA-256 digest after download and before every use. A corrupt, tampered, or attacker-planted model is rejected. If a cached model fails the check, Kiro deletes it and re-downloads from the official CDN automatically. No action is needed.
• On Unix systems, the model directory is restricted to owner-only access (mode 0700) as defense in depth.

> /voice
Recording, press ENTER when done... 2.1sec ░░░████████

"Fix the bug in the authentication module"

> Analyzing the authentication module...

1. Type the start of your prompt: How do I
2. Hold Space and speak: "implement pagination in the users API"
3. Release Space to stop and transcribe
4. The submitted prompt reads: "How do I implement pagination in the users API"

• Requires a physical microphone, or a remote voice server for cloud desktops.
• Local transcription is CPU-bound. The small model is more accurate but slower, and there is no GPU acceleration yet.
• Optimized for English by default; multilingual support is limited.
• No wake word activation.

• Chat: the interactive session that voice input feeds into
• Terminal UI: the interface that renders the live recording indicator
• In-session settings: adjusting settings during a session