Voice Input — Soniox STT for VSCode

Push-to-talk voice-to-text that works in any input across VSCode — code editor, GitHub Copilot Chat, Claude Code chat, terminals, settings forms. Audio is captured natively (no flaky webview microphone), transcribed via the Soniox STT API, and pasted into the focused element.

Hebrew is the default language. The full UI is bilingual (he / en) with native RTL support.

Features

• Background recording. Toggles recording from anywhere (default: Alt+M on Linux/Win, Ctrl+Alt+M on macOS) — customizable in sidebar, shows your configured shortcut dynamically.
• Universal injection. Editor → cursor insertion. Chat webviews / panels → clipboard + Ctrl+V simulated via ydotool. Works around VSCode's webview microphone block.
• Audio device selection. Voice Input: Select Audio Device lists every available microphone on the current platform and lets you pick one with a QuickPick. The selection is saved to voiceInput.audioDevice. On Linux the list updates instantly when a USB mic is plugged or unplugged; on macOS/Windows it refreshes automatically within seconds.
• No-device guard. If no audio input source is detected when you try to start a recording, the extension blocks the attempt and offers a Select Device button instead of surfacing a cryptic recorder error.
• Speech history. Every transcription is saved with timestamp + language. One-click copy or delete per entry. Configurable TTL (1 day / 7 days / 30 days / forever).
• In-panel settings. Speech language, UI language, history TTL, Soniox model, recording shortcut, API key — all editable from the sidebar without leaving the editor.
• Bilingual UI. Hebrew (default) and English with automatic RTL/LTR layout.
• Secure key storage. SONIOX_API_KEY is held in VSCode SecretStorage, never in settings.json.
• Append-only. Never overwrites a selection, never auto-submits a chat message.
• Diagnostics built-in. Output channel logs every step; one command dumps environment + tool availability.

Keyboard shortcuts

The shortcut works from any focus — editor, chat input, terminal, and the Voice Input sidebar panel — and never moves your view. The transcription lands at the cursor (editor) or is pasted via simulated Ctrl+V (chat / other inputs).

Trigger on key-up. When the recording shortcut is pressed while the Voice Input sidebar panel has focus, the toggle fires on key release (not keydown). This prevents accidental double-triggers from key-repeat and matches the push-to-talk feel of the mic button. Your currently configured shortcut is displayed in the sidebar.

To change a binding by hand: Ctrl+K Ctrl+S → search voiceInput.toggleRecording → click the pencil → press your new combo.

Platform support

The extension auto-detects the active platform and chooses the right backend at runtime — no per-platform config needed.

Requirements

One-time setup on macOS

brew install ffmpeg

That's it. osascript and pbcopy come with macOS. After installing the extension, grant VSCode microphone access in System Settings → Privacy & Security → Microphone (the OS will prompt on first recording).

The recording shortcut on macOS is Ctrl+Alt+M (Ctrl+Option+M). Paste uses Cmd+V automatically.

One-time setup on Linux Wayland (Ubuntu / Debian)

sudo apt install ffmpeg ydotool wl-clipboard

sudo tee /etc/systemd/system/ydotoold.service >/dev/null <<EOF
[Unit]
Description=ydotool daemon

[Service]
Type=simple
ExecStart=/usr/bin/ydotoold --socket-path=/tmp/.ydotool_socket --socket-own=$(id -u):$(id -g)
Restart=always

[Install]
WantedBy=multi-user.target
EOF

sudo systemctl daemon-reload
sudo systemctl enable --now ydotoold.service

The extension reads /tmp/.ydotool_socket automatically.

One-time setup on Linux X11

sudo apt install ffmpeg xdotool xclip

One-time setup on Windows

winget install Gyan.FFmpeg

powershell and clip.exe ship with Windows. After install, restart your shell so ffmpeg is on PATH. The extension auto-detects the first DirectShow audio device; override with voiceInput.audioDevice in settings (e.g. Microphone (Realtek)). Find your device name with:

ffmpeg -hide_banner -list_devices true -f dshow -i dummy

Quick start

1. Install the extension — code --install-extension voice-input-*.vsix.
2. Reload VSCode (Developer: Reload Window).
3. Open the Voice Input view in the activity bar (purple mic icon).
4. Click Set Soniox API key in the Settings section and paste your key.
5. Click into any chat input or editor.
6. Press Alt+M (Linux/Win) or Ctrl+Alt+M (macOS) to start recording — the status bar turns red.
7. Press the shortcut again — the transcript is pasted at the cursor and saved to history.
8. The Voice Input view shows the full history with copy / delete buttons.

Settings

Configurable from both the in-panel Settings section (collapsible) and settings.json:

To pick a device interactively run Voice Input: Select Audio Device from the Command Palette — it enumerates all available inputs and writes the chosen id to voiceInput.audioDevice automatically.

auto injection: text-file tab → editor cursor; everything else → clipboard + simulated Ctrl+V.

Commands

All available from the Command Palette (Ctrl+Shift+P):

Show Diagnostics opens the Output panel and logs: extension version, session type (Wayland / X11), availability of ffmpeg / ydotool / wl-copy, and ydotool socket state. Use it whenever paste or recording misbehaves.

Why two paths for injection?

VSCode webviews (used by Claude Code chat, etc.) are sandboxed and reject getUserMedia and any programmatic text injection from other extensions. The workaround is OS-level: capture audio via ffmpeg from the PulseAudio default source, then simulate a real Ctrl+V keystroke against the focused window so the chat input handles paste like any other clipboard event. Editors get the cleaner path via vscode.TextEditor.edit.

Troubleshooting

Changelog

See CHANGELOG.md for the full history.

v1.0.7

• Feat: Audio device selector is now embedded in the extension sidebar Settings section — a dropdown lists all detected microphones with a ↺ Scan button to refresh the list without leaving the panel. The same setting is also editable in the standard VSCode Settings UI (voiceInput.audioDevice).

v1.0.6

• Feat: Voice Input: Select Audio Device command — enumerates all microphones on the current platform (PulseAudio/PipeWire on Linux, AVFoundation on macOS, DirectShow on Windows) and lets you pick one from a QuickPick. Selection is saved to voiceInput.audioDevice.
• Feat: Dynamic device list refresh. On Linux the extension watches /dev/snd/ with a filesystem watcher so plug/unplug events are reflected immediately. On macOS/Windows the cache has a 5-second TTL so stale data is never shown for more than a few seconds.
• Feat: No-device guard — if no audio input is detected when starting a recording, the attempt is blocked with a clear error and a Select Device shortcut instead of a cryptic recorder failure message.

v1.0.5

• Fix: Added missing icon property to the microphone webview panel — resolves VSCode manifest validation warning.

v1.0.4

• Feat: The hint text below the mic button ("or press … to toggle") now displays your currently configured keyboard shortcut dynamically. Changing the binding via the VSCode Keyboard Shortcuts editor is reflected immediately — no hardcoded Alt+M in the UI.
• macOS shows Ctrl+Alt+M by default; Linux/Win show Alt+M.

v1.0.3

• Fix: Settings dropdowns no longer overflow the sidebar panel width at narrow sizes.

v1.0.2

• Fix: Alt+M now works when the Voice Input sidebar panel has keyboard focus. Previously the keydown event was silently swallowed by the webview browser context with no handler, so the shortcut had no effect. A keydown/keyup listener pair is now registered in the webview — keydown suppresses the default action, keyup fires the toggle.
• Behaviour: When triggered from within the webview, the toggle fires on key release (keyup) rather than keydown, avoiding repeated triggers from key-repeat.

v1.0.1

• Added MIT LICENSE file and repository field to package.json.

v0.3.6

• Platform install scripts (scripts/install-linux.sh, install-mac.sh, install-windows.ps1).
• Auto dependency check on activation — surfaces a notification if ffmpeg, ydotool, or clipboard tools are missing.

v0.3.5

• Initial release: push-to-talk recording, Soniox STT, universal injection, bilingual UI (he/en), speech history, secure API key storage, sidebar settings panel.

License

MIT.