dsh-voice-scribe：DeepSeek Harness で音声入力（書き起こし）｜導入ガイド - DeepSeek Harness (dsh) プラグインの使い方・インストール

MIT ライセンスの DSH ネイティブ音声書き起こしプラグイン。DeepSeek Harness の Web 入力欄向けで、SenseVoice によるオフライン優先、ブラウザ／クラウド ASR フォールバック、および DSH モデルを再利用する任意の推敲ステップを備えます。

DSH 統合

ネイティブ実行

作者による申告

安全性監査

未監査

最終検証日

2026-09-06

ライセンス

MIT

01どんなタスクに使えるのか？

Dictate text hands-free into the DeepSeek Harness web composer
Transcribed speech inserted at the composer cursor as editable text
DeepSeek Harness (dsh) web users who want voice-to-text input
fa7be16aa833…
Transcribe speech with the best available engine (offline-first)
Text via SenseVoice offline (audio never leaves the machine) or browser / cloud ASR fallback
Privacy-conscious dsh users and those behind restricted networks
fa7be16aa833…

インストール手順

導入成功の確認

作者は未記載

DSH 統合ネイティブ実行

Installed via the official `dsh plugin --profile web add` command; injects voice dictation into the dsh web composer

Alt hotkey & microphone button dictation
Speech via microphone (tap/hold Alt, or click the 🎤 button)→Transcribed text inserted at the composer cursor
fa7be16aa833…
Multi-engine recognition (offline-first)
Microphone audio→Text via SenseVoice offline / browser Web Speech / cloud ASR service chain
First use auto-downloads the SenseVoice model (~230MB, one-time)
fa7be16aa833…
Hold-to-talk mode
Hold the hotkey to record→Auto-transcribe on release
fa7be16aa833…
Hot-word replacement table
Edit $DSH_HOME/voice/hot.txt→Corrected transcription (literal + regex rules) applied to all engines
fa7be16aa833…
Optional LLM polish (reuses the DSH model)
Raw transcript + custom polish prompt→Cleaned text via local rule pre-polish then the DSH LLM
Polish prompt and transcript are sent to the DSH model
fa7be16aa833…

向いている用途

不向きな用途

Unofficial community plugin, not affiliated with DeepSeek / 深度求索; read SECURITY.md before use.
Browser Web Speech relies on external services (Chrome is blocked in mainland China, Edge Stable has known regressions); local recognition is the default.

Unofficial community plugin, not affiliated with DeepSeek / 深度求索; read SECURITY.md before use.
Browser Web Speech relies on external services (Chrome is blocked in mainland China, Edge Stable has known regressions); local recognition is the default.
Maximum recording length is ~4 minutes on the local engine and 10 minutes on cloud ASR; recording auto-stops at the limit.
The cloud ASR service chain requires configuring an API key in settings; the key is stored server-side only.