dsh-vidi

English | 中文

A video tool for DeepSeek Harness: watch live streams, videos and podcasts with subtitles in any language, and let DeepSeek take frames, cut clips, download and summarise for you.

What it does

• Opens anything: YouTube, Bilibili and other sites yt-dlp supports, Xiaoyuzhou podcast episodes, m3u8 / rtmp / srt streams, local files (drop them in), or a stream pushed from OBS.
• Subtitles start by themselves when a video opens; pause and resume any time. One or two languages (a main line and a second line); the site's own subtitles can be shown too, or used as a reference.
• Two paces: real-time shows lines as they are heard; polished arrives a little later with misheard words fixed from context.
• The subtitle model looks at the video: polishing uses the page's title, description and comments, and — when the chosen model reads images — a frame of that stretch. It also keeps a glossary of names and terms for the video.
• Chat with DeepSeek beside the video: grab this moment, cut this part, summarise, download the original — all in one conversation, with files saved in the video's folder.
• Restream: push the video with its subtitles drawn into the picture to Bilibili, Douyin, Kuaishou, WeChat Channels, Douyu, Huya, YouTube, Twitch or any RTMP/SRT address — several at once; each platform connects and reconnects on its own.
• Export: SRT / VTT subtitles, transcripts by line or paragraph, clips with subtitle tracks.
• English and Chinese interface; works on phones.

Install

In DSH, open Plugins → Install third-party plugin and enter dsh-vidi. A Video button appears in the sidebar; pasting a video link in a chat also offers to open it.

Needs DSH 0.2.2 or later.

Dependencies

Tool	For	Install (macOS)
ffmpeg	playing, recording, frames, clips	brew install ffmpeg
yt-dlp	site videos	brew install yt-dlp
whisper.cpp (optional)	transcription on this computer	brew install whisper-cpp, then put ggml-large-v3-turbo-q5_0.bin in ~/.dsh/live/models/

Without whisper.cpp, pick an online service as the transcription model in settings.

Two models

Set separately under the gear:

• Transcription: speech to text. Whisper on this computer by default; or OpenAI, Google Gemini, Groq, SiliconFlow, Zhipu GLM.
• Subtitle model: polishing, translation, glossary. DeepSeek by default; the list shows every model configured in DSH and marks the ones that read images.

Online transcription uses the model providers you have set up in DSH (Models page) — OpenAI, Google, Groq, SiliconFlow, Zhipu GLM: their key and address work as they are, whatever the key is called. A service DSH doesn't have yet can get a key right on the video page; it is checked, then saved in DSH.

Restreaming

In the player, open Restream, add a platform (paste the server address and stream key from the platform's own streaming tool — e.g. Bilibili 直播姬, Douyin 直播伴侣, WeChat Channels Assistant), tick the ones to use and start. It starts from where you are watching, with the subtitles you see; for a live stream that is a few seconds behind live, so the lines are ready. Stream keys stay on this computer.

Streaming from OBS

Open the video page: the empty page shows the server address and stream key. In OBS choose the "Custom" service and paste them. Default ports: RTMP 1935, SRT 9000 (changeable in the plugin settings).

Where things are saved

Recordings, subtitles, frames and clips live in ~/.dsh/live/jobs/<id>/ (changeable in the plugin settings).

Known limits

• Xiaoyuzhou video episodes only offer a one-minute preview on the web (the full video needs the app and a login), so they play as the full audio with the cover.
• Site videos that need a login use a browser's saved login (Settings → Site login).

License

PolyForm Noncommercial 1.0.0: use, change and share it freely for anything except commercial purposes. No copyleft: your own changes can stay under whatever terms you like, as long as the use is noncommercial.