Parakeet voice input for DeepSeek Harness

This plugin adds a local microphone button to the DeepSeek Harness composer. Audio stays on the user's computer. The plugin never sends a message by itself.

Quick start

1. Install and start DSH Desktop once.

2. Download this repository (or open its folder in PowerShell).

3. Run the installer:

   .\install.ps1
   

4. Install the model using Install the model.

5. Restart DSH Desktop.

6. The microphone button appears in the composer. You can click it, or press Ctrl+E once, to start recording. Press the same shortcut once more to stop.

7. Review the transcript in the draft and press Send.

You do not need to click the button before using the shortcut, and you do not need to hold either control. The shortcut calls the same local recording action as the button.

The installer uses the official dsh plugin command when it is available. DSH Desktop users normally do not have that command on PATH, so the same script falls back to the Desktop profile at $DSH_HOME/profiles/desktop, registers the bundle layer, and removes the conflicting built-in voice-input row. It does not write to the DSH application directory.

Shortcut

The default two-key shortcut is Ctrl+E on Windows/Linux and Command+E on macOS. Press it once to start recording and once again to finish.

Right-click the microphone to choose another shortcut. The new shortcut must include Ctrl, Alt, Shift, or Command, and is saved in the browser. You can also open Settings → Plugins → Parakeet Voice Input and change the shortcut there. That settings page is the recommended way to configure it; clicking the microphone is only for starting or stopping a recording.

Install the model

The Parakeet model runs offline. Audio stays on the computer. For safe installation, provide a trusted download URL and the model publisher's SHA-256 checksum:

.\install-model.ps1 `
  -Url 'https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3/resolve/main/parakeet-tdt-0.6b-v3.q8_0.gguf?download=true' `
  -Sha256 '<64-character SHA-256 checksum>'

If the model is missing, DSH still starts normally and the microphone remains disabled. The installer verifies the checksum before replacing the active file.

How recording works

• The first click starts the local companion and microphone capture.
• The second click stops recording.
• The offline model receives one local WAV file and returns the final transcript.
• The transcript is inserted into the draft; sending remains a separate action.
• The companion stops after recording ends or when the plugin unloads.

Offline models do not show interim words. The final transcript appears after the second click.

Update or remove

Update the installed plugin and then restart DSH:

dsh plugin --profile desktop update @local/dsh-parakeet-voice-input

The update preserves the model under $DSH_HOME/models. Remove the plugin with:

dsh plugin --profile desktop remove @local/dsh-parakeet-voice-input

If dsh is not on PATH, run install.ps1 again to update. To remove the Desktop installation cleanly, use:

corepack pnpm --dir "$env:USERPROFILE\.dsh\profiles\desktop" remove @local/dsh-parakeet-voice-input

Then remove the two Parakeet rows from that profile's cordis.patch.yml and restart DSH Desktop. The model file is kept so reinstalling does not require a new download; delete $DSH_HOME\models\parakeet-tdt-0.6b-v3 only if you also want to remove the offline model.

Troubleshooting

If the microphone stops immediately:

1. Restart DSH after installing or updating the plugin.
2. Check that the model exists under $DSH_HOME/models/parakeet-tdt-0.6b-v3.
3. Check the browser microphone permission.
4. Make sure port 8080 is not used by another process.
5. Click once, wait a second, and then speak.

If the button is missing after installation, restart DSH Desktop once more. The installer must have access to the Desktop profile; running pnpm add in this repository alone does not activate a DSH bundle.

If the button is visible but Ctrl+E does nothing, click inside the chat once so the WebView has focus, then press Ctrl+E. The listener is installed on the whole chat window and does not require the microphone button to be focused.

The microphone permission prompt is expected. The chat is rendered in a local browser/WebView, and browsers require explicit getUserMedia permission before any page can read a microphone. Audio is sent only to the local Parakeet service at 127.0.0.1; it is not uploaded to a speech API.

Requirements

• DeepSeek Harness Web profile;
• Windows x64 (the package includes a Windows companion);
• a browser with microphone and Web Audio support.

Verify the package before publishing:

node scripts/verify-package.mjs