dsh-local-transcribe

A DeepSeek Harness (DSH) plugin for fully-local speech-to-text powered by whisper.cpp. Everything runs on your machine — nothing is uploaded.

• 🎛 Visual panel (sidebar icon): drag & drop or browse local audio files, pick several and merge them into one recording, transcribe with live progress, preview the full text with a per-segment timeline, play back the merged audio, and edit the transcript in place before saving.
• 🤖 Agent tool transcribe_audio: every agent session can transcribe one file or merge several files (in order) into one continuous transcript.
• 🈶 Supports m4a / mp3 / wav / flac / aac / ogg / opus / wma / m4b / amr / aiff / aif / caf and video files (mp4 / mov / mkv / avi / webm / m4v / ts / 3gp / mpg / mpeg).
• 🔒 Panel–host communication is loopback-only (127.0.0.1 / localhost), owned by the plugin context.

Requirements

1. A running DSH (≥ 0.2.2-rc.1) with a web profile.
2. ffmpeg / ffprobe on PATH (brew install ffmpeg on macOS).
3. A whisper.cpp build with the whisper-cli executable, plus a ggml model:

   git clone https://github.com/ggml-org/whisper.cpp
   cd whisper.cpp && cmake -B build -DCMAKE_BUILD_TYPE=Release && cmake --build build -j
   # model, e.g. Chinese-friendly medium (~1.5 GB):
   curl -L -o ggml-medium.bin https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-medium.bin
   

Install

Clone the repository and install the bundle from the Plugin Manager, or install it directly:

git clone https://github.com/bizbin/dsh-local-transcribe
# In DSH, ask the agent (or use the Plugin Manager UI) to:
#   install_bundle <absolute path to dsh-local-transcribe>

Configure

In your profile's cordis.patch.yml, set the local-transcribe row's config:

- id: local-transcribe
  config:
    whisperCli: /path/to/whisper.cpp/build/bin/whisper-cli
    model: /path/to/ggml-medium.bin
    # ffmpeg: ffmpeg            # command on PATH by default
    # ffprobe: ffprobe
    # language: zh              # default transcription language ('auto' to detect)
    # prompt: '以下是普通话的句子。'
    # threads: 10

Key	Default	Meaning
whisperCli	whisper-cli (PATH)	whisper-cli executable
model	(required)	absolute path of the ggml model file
ffmpeg / ffprobe	ffmpeg / ffprobe	commands on PATH or absolute paths
language	zh	default language code
prompt	以下是普通话的句子。	initial whisper prompt
threads	10	whisper-cli thread count

Panel

Open the waveform icon in the sidebar. Then either drag audio files onto the panel (they are uploaded to ~/.dsh/storages/local-transcribe-uploads/, 500 MB per file) or browse the filesystem. Order the selected files, pick an output directory, language and prompt, and press Transcribe. While running you see the phase and percentage; when done you get the transcript, a segment timeline, and merged-audio playback. Edit the text and press Save edits — the plugin rewrites <name>_原始转写.txt and regenerates <name>_转写稿.md.

Outputs (next to the first input, or output_dir):

• <name>_原始转写.txt — plain text
• <name>_带时间轴.json — segments with millisecond timestamps
• <name>_转写稿.md — markdown, one heading per minute

Agent tool

transcribe_audio — one file via path, or several merged in order via paths; language, prompt, output_dir, output_name, markdown, keep_wav options. The tool returns the full text, the segment timeline, and the output file paths.

Security notes

• Transcription uses only local binaries; audio never leaves the machine.
• The panel's HTTP routes (/local-transcribe/rpc, /local-transcribe/upload, /local-transcribe/play/*) accept requests from loopback only, and are removed when the plugin unloads.

License

MIT

---

dsh-local-transcribe（中文说明）

基于 whisper.cpp 的 DeepSeek Harness（DSH）插件，完全本地完成语音转文本，音频不出本机。

• 可视化面板（侧边栏波形图标）：拖拽或浏览选择音频文件，多个文件按顺序合并为一段整体识别，实时显示进度；完成后可预览全文与分段时间轴、试听合并音频、就地编辑文本并保存（重写 txt 并重新生成 markdown）。
• 智能体工具 transcribe_audio：任意会话可调用，支持单个文件或 paths 数组合并识别。
• 支持 m4a / mp3 / wav / flac / aac / ogg / opus / wma / m4b / amr / aiff / aif / caf 等音频与 mp4 / mov / mkv 等视频。
• 面板与宿主通过仅限环回地址的 HTTP 路由通信，随插件卸载自动注销。

依赖：DSH ≥ 0.2.2-rc.1；ffmpeg/ffprobe；whisper.cpp 编译出的 whisper-cli 与一个 ggml 模型（下载命令见上文）。安装后在本 profile 的 cordis.patch.yml 中配置 whisperCli 与 model 两个路径即可使用。