dsh-dictation

When composing long prompts in DeepSeek Harness, typing can become a real bottleneck. Many voice-input solutions either auto-send the moment you stop speaking or refuse to let you edit the transcript before it goes out. dsh-dictation takes a different approach: it writes the recognized text into the editable draft box so you can proofread, fix typos, and adjust wording before deciding to hit send. The microphone button sits right next to the send button; you can also toggle recording with Ctrl+Shift+D, and pressing Escape during a local recording cancels the session without disturbing your existing draft.

The plugin ships with no speech model bundled. Instead, a dedicated Settings page lets you pick, download, and uninstall models on demand. Four sources are currently available: SenseVoice Small (roughly 228 MB, covering Mandarin, Cantonese, English, Japanese, and Korean), Paraformer Small (roughly 78 MB, fast Mandarin and English recognition), Nemotron 3.5 ASR (roughly 712 MB, multilingual local recognition with automatic language detection), and a Beta global-dictation mode that reuses an already-installed Codex Desktop. Every local model runs entirely on your machine, and the Codex mode never reads account credentials or personal data.

If you regularly type long inputs in DeepSeek Harness, care about keeping speech data local, or need multilingual dictation, dsh-dictation is a lightweight and controllable add-on. It is open-sourced under the MIT license and validated against the latest stable release of the Harness.