dsh-local-voice-dictation

Created by LionGateOS

Author: Antonio Vidal antoniovidaljr@gmail.com

Local voice for DeepSeek Harness. Dictate with your microphone through local speech-to-text, then listen to finalized assistant responses through local Kokoro text-to-speech.

• Local STT dictation — microphone audio stays on your local voice path.
• Local Kokoro TTS — Speak controls on finalized assistant responses.
• No auto-send — dictated text is inserted into the draft for review.
• Playback control — visible generation state and Stop speaking support.

Independent community plugin; not an official DeepSeek project or upstream PR. Licensed under MIT.

Installation

Installation is currently manual. See INSTALL.md for the current setup procedure and requirements.

What it does

Local voice dictation for DeepSeek Harness web sessions — no cloud STT, no auto-send:

1. A Dictate button in the composer records microphone audio in the browser (WebRTC getUserMedia).
2. The audio is sent to the same-origin proxy route POST /voice-proxy/transcribe (base64 JSON payload, capped at about 22 million base64 characters, approximately 16 MB of decoded audio).
3. The proxy forwards the audio to a local STT engine on the same machine and receives the transcript back.
4. The transcript is inserted into the composer draft.

It never auto-sends messages. The transcript always lands in the draft; you review it and send it yourself.

Speak

Each finalized assistant response gets a speaker control.

The Speak flow uses local Kokoro text-to-speech:

1. Click the speaker attached to an assistant response.
2. The button immediately shows Generating....
3. The response text is sent through the same-origin route POST /voice-proxy/speech.
4. The proxy forwards the text to the local Kokoro-compatible voice engine.
5. When playback starts, the control becomes Stop speaking.
6. Only one generated audio stream is allowed at a time, preventing delayed duplicate clicks from producing overlapping voices.

The currently tested Kokoro endpoint is:

POST http://127.0.0.1:8768/v1/audio/speech

The assistant action is mounted using DeepSeek Harness's conversation.chat.assistant-actions slot.

The local flow

Dictate:
browser mic  →  same-origin /voice-proxy/transcribe  →  local engine (127.0.0.1)  →  transcript  →  composer draft

Speak:
assistant response  →  same-origin /voice-proxy/speech  →  local Kokoro TTS  →  WAV  →  browser playback

How it's built

Two halves in one small package (~330 lines of JS):

File	Role
index.js	Node half — registers the same-origin /voice-proxy/transcribe and /voice-proxy/speech routes
client.js	Browser half — Dictate button, mic capture, transcript insertion, and per-assistant-message Speak controls
package.json	Manifest: @local/voice-dictation, ESM, node main + ./client export

Required engine interface

The add-on is engine-agnostic: it works with any local engine that exposes

• GET /health — a status check, and
• POST /transcribe — a multipart audio upload returning JSON with a text field and an optional language field.

The current local LionGate Voice Engine (faster-whisper STT) is one compatible local engine implementation; the prototype's proxy is pointed at its /transcribe endpoint.

Mounting

Mount the plugin in the DeepSeek Harness web browser roster:

packages/bundle/web-app/cordis.patch.yml

Add this browser plugin row:

- id: voice-dictation
  name: '@local/voice-dictation'

Do not mount the plugin only inside an agent preset. Agent presets do not add the client plugin to the web browser boot graph.

Note: an earlier prototype packaged this as a full copied "Voice Mode" agent preset. That approach turned out to be unstable (default-preset drift, source living only in gitignored node_modules) and is not recommended.

Requirements

• A DeepSeek Harness web UI (developer preview)
• A local STT engine reachable on 127.0.0.1
• A Kokoro-compatible TTS endpoint for Speak
• Browser microphone permission

If the engine is down, the proxy returns HTTP 502 with a clear error and the rest of Harness is unaffected — this behavior is code-backed, and a direct engine-down test is still on the roadmap (needs direct testing to be fully validated).

Status

• ✅ Dictation round-trip demonstrated: record → proxy → engine → draft
• ✅ No auto-send; transcript always lands in the draft
• ✅ Assistant-response Speak button
• ✅ Local Kokoro TTS through /voice-proxy/speech
• ✅ Visible Generating... state while speech is being prepared
• ✅ Duplicate generation clicks are blocked
• ✅ Only one plugin-generated audio stream plays at a time
• ✅ Stop speaking immediately stops active playback
• ⚠️ Installation is manual for now (see INSTALL.md)