JKinco Listen · 筑听（开源本地版）

JKinco Listen (Open Edition)

简体中文

A local-first AI meeting-minutes workbench. Recording → transcript → scene detection → structured minutes → DOCX/PDF, entirely on your own machine. No API key, no cloud call.

---

The constraint this is built for

Some meetings cannot be uploaded. Not “should not” — cannot.

A construction site review decides who is accountable for a defect. An interview record holds a candidate’s personal data. A client visit contains pricing. In regulated industries these recordings are part of a legal archive, and sending them to a third-party server is not a policy preference you can toggle.

Cloud meeting assistants are architecturally unable to serve that case. Their value comes from models running on their infrastructure, so the audio has to get there. That is not a gap in their product; it is what their product is.

So the people in those meetings do the only thing left: they type the notes by hand.

This project is for them. Everything runs on the machine in the room.

What comes out

Summary, agenda flow, conclusions, and action items — with the full recording-to-export pipeline on the right:

Measured, not claimed

Ten noisy Chinese meeting clips, 36 terms a general model tends to mishear:

	Terms recognised
Local ASR alone	31 / 36
+ domain correction	36 / 36

Reproduce it in one command — see the benchmark page for the method, the per-clip results, and the three bugs writing that benchmark found in this project’s own code.

Quick start

git clone https://github.com/wenxuanzhang1209-cyber/jkinco-listen-open.git
cd jkinco-listen-open
cp .env.example .env
docker compose up -d --build

Open http://localhost:8080.

⚠️ Change the password before exposing this to a network. .env.example ships with JKINCO_AUTH=admin:123456 so a local trial works immediately, and the container binds 0.0.0.0 — anyone who can reach the port can sign in with it.

Want to see the interface before downloading 2 GB of models? Set JKINCO_DEMO_DATA=1 and two example meetings appear, with export, search, and templates all usable.

How it compares

	JKinco Listen	Cloud note-takers	Raw Whisper
Audio leaves your machine	Never	Yes	Never
API key required	No	Yes	No
Output	Formatted DOCX/PDF	Transcript + summary	Raw transcript
Meeting-type awareness	5 scenes, evidence-gated	One format	None

If you only need a transcript, use Whisper directly — it is simpler. This exists for the part after the transcript: turning it into a document somebody can file.

Under the hood

• 🎙 Local speech recognition — FunASR paraformer-zh
• 🧠 Local LLM — Ollama (qwen2.5:7b-instruct by default) or any OpenAI-compatible endpoint
• 🏗 Evidence-gated scene detection — rules decide first; the model may downgrade a classification but never promote one, because the wrong template produces a document asserting the wrong legal responsibility
• 📄 Original-layout DOCX/PDF export
• 🔐 No API key, no cloud, no telemetry — enforced in CI, not asserted in a README
• 🐳 One-command Docker deploy

Documentation

• Project home (README)
• Benchmark
• Architecture
• Model guide
• Roadmap
• About the Open Edition

Community

• Discussions
• Issues

If this saved you from typing up one more set of meeting minutes, please ⭐ the repository.

MIT License © 2026 JKinco

---

_{JKinco — local-first tools for work whose data cannot leave the building · Listen · Slides · Skills Lab · Life Hub · Tools}

This site is open source. Improve this page.