JKinco Listen (Open Edition)

A local-first AI workbench that turns meeting recordings into finished minutes.

Recording → transcript → scene detection → structured minutes → DOCX/PDF, running entirely on your own machine. No API keys. No cloud calls. Your audio never leaves your computer.

简体中文 · Docs site · Architecture · Model guide · Roadmap · Changelog

---

Why this exists

Almost every meeting-notes product sends your audio to someone else's servers: upload the recording, pay per minute, trust a third party with the contents.

For a construction site review, a customer visit, or a job interview, that trade is often unacceptable — so people fall back to typing up notes by hand.

JKinco Listen moves the whole pipeline onto your machine:

• Local speech recognition — Chinese meeting-grade ASR with voice-activity detection, punctuation restoration, and domain-term correction.
• Local LLM — scene detection, minutes generation, quality review, and meeting Q&A all run against a model you host (Ollama, or any OpenAI-compatible endpoint).
• Local template engine — five meeting types exported to DOCX/PDF with the original corporate layout preserved.
• Local history — meetings, transcripts, minutes, and human edits stay in a SQLite file you own.

One command to deploy. Audio in, minutes out. Nothing leaves the building.

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 — it is off the table.

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 exists for them. Everything runs on the machine in the room.

_{有些会议不是"最好别上传"，是不能上传。工程例会决定责任归属、面试记录含个人信息、 客户拜访含报价——在受监管的行业里，这些录音属于法律档案，传给第三方服务器不是一个可以 开关的选项。云端会议助手在架构上就服务不了这类场景：它们的价值来自跑在自己机房里的模型， 音频必须传过去。于是这些会议的与会者只能手打纪要。这个项目就是为他们做的。}

How it compares

	JKinco Listen (Open)	Cloud note-takers	Raw Whisper / FunASR
Audio leaves your machine	Never	Yes	Never
API key / subscription	None	Required	None
Works fully offline	Yes, after first model download	No	Yes
Output	Formatted DOCX/PDF minutes	Transcript + summary	Raw transcript
Domain accuracy	Hotword lexicon + post-correction	Generic	Generic
Meeting-type awareness	5 scenes, evidence-gated	One format	None
Self-hosted, MIT licensed	Yes	No	Library only

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

Measured, not claimed

Full results and methodology: docs/BENCHMARK.md

Ten noisy Chinese meeting clips, 36 domain terms that a general model tends to get wrong (construction, HR, and commercial vocabulary), run through the local paraformer-zh:

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

Every one of the six failures was a construction term:

航站记录 ← 旁站记录      微大工程 ← 危大工程
石滩落度 ← 时坍落度      同条件是块 ← 同条件试块

The correction table is built from what the model actually gets wrong, not from what a human might mistype — the first thirteen entries were written by imagining plausible errors, and when this benchmark ran, not one of them fired. Reproduce it yourself — one command, no arguments:

pip install funasr torch torchaudio numpy
python3 scripts/benchmark_domain_accuracy.py

It synthesises the clips, degrades them with noise and tempo shift, runs the local model, and prints the same table. The audio is synthetic and therefore cleaner than a real meeting room — treat the number as an upper bound. Its value is comparability: change the model or the lexicon, run it again, and the delta is trustworthy.

_{十段加噪中文会议音频、36 个通用模型容易听错的领域词，过本地 paraformer-zh： 仅本地识别 31/36，加上领域纠错 36/36。六处失误全在工程术语上。 纠错表是照着模型真正会犯的错写的——最初那 13 条是想象出来的，跑基准时一条都没命中。}

Highlights

Two local model stacks	FunASR paraformer-zh for speech + any Ollama / OpenAI-compatible model for text
Evidence-gated scene detection	Rule-based gate first, model second — the model cannot force a construction template onto a meeting that lacks the evidence
Original-layout export	DOCX/PDF that match the real corporate templates, not a generic markdown dump
Privacy by construction	No API keys, no telemetry, no outbound model calls; CI fails the build if cloud-model traces appear in the repo
One-command deploy	docker compose up brings up web, backend, Ollama, and model pull
961 automated tests	Plus security headers, rate limiting, injection defenses, and an audit log
Local knowledge base	Search past meetings and ask questions across them, without anything leaving the machine

Screenshots

_{Full-resolution MP4: docs/demo.mp4}

This is what comes out — summary, agenda flow, conclusions, and action items, with the full recording-to-export pipeline on the right:

More screens (click to expand)

Sign-in — a local account, no cloud registration

Workspace — upload, live recording, or read straight from a recorder; seven scene tabs

History — per-scene stats, full-text search, reopen and export any past meeting

Quick start (Docker)

Requires Docker. 16 GB RAM recommended; 8 GB works but is slower.

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.

Set a password before you expose this to a network. .env.example ships with JKINCO_AUTH=admin:123456 so that a local trial works immediately. The container listens on 0.0.0.0, so anyone who can reach the port can sign in with those credentials. Change JKINCO_AUTH in .env before binding it to anything other than localhost.

The first start automatically:

1. downloads the local ASR model (~1–2 GB; fully offline afterwards);
2. pulls the local LLM qwen2.5:7b-instruct via Ollama;
3. builds the web UI and starts the backend.

Want a different model? OLLAMA_MODEL=qwen2.5:14b docker compose up -d — nothing else changes.

NVIDIA GPU:

docker compose -f docker-compose.yml -f docker-compose.gpu.yml up -d

Apple Silicon needs no configuration; Ollama uses Metal automatically.

Try it before downloading any model:

docker compose -f docker-compose.yml -f docker-compose.demo.yml up -d

Demo mode seeds two example minutes so the UI, exports, and history search are all usable without a model.

Manual setup

One command, for people who would rather not use Docker:

bash scripts/install.sh        # Ollama model + Python deps + frontend
bash scripts/start.sh

Or step by step:

# 1. Local LLM (pick one)
brew install ollama && ollama pull qwen2.5:7b-instruct   # macOS
curl -fsSL https://ollama.com/install.sh | sh             # Linux

# 2. Backend
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env

# 3. Frontend
cd frontend && npm ci && npm run build && cd ..

# 4. Run
uvicorn backend.main:app --host 0.0.0.0 --port 8080

Larger model: OLLAMA_MODEL=qwen2.5:14b bash scripts/install.sh

Model guide

Stage	Default	Alternatives	Notes
Speech recognition	FunASR paraformer-zh	fsmn-vad + ct-punc	Best value for Chinese meetings; runs on CPU
Scene review	Ollama qwen2.5:7b-instruct	qwen2.5:14b, qwen3:8b	The evidence gate has the final say, so a wrong model answer cannot flip the result
Minutes generation	Ollama qwen2.5:7b-instruct	Larger models	Chunk → extract in parallel → merge → quality review
Meeting Q&A	Same	Same	Local retrieval and generation; history stays on the machine

Hardware and quantization guidance: docs/LOCAL_MODELS.en.md.

Architecture

flowchart LR
    U[Browser · React + TS] -->|upload / live recording| API[FastAPI backend]
    API --> ASR[FunASR local recognition]
    ASR --> LEX[Domain lexicon correction]
    LEX --> CLS[Scene detection<br/>evidence gate + model review]
    CLS --> LLM[Local LLM via Ollama<br/>OpenAI-compatible API]
    LLM --> MIN[Structured minutes + quality review]
    MIN --> EXP[DOCX / PDF export]
    MIN --> DB[(Local history · SQLite)]
    DB --> QA[Ask JKinco · local meeting Q&A]

Loading

Module boundaries, data flow, and the security perimeter: docs/ARCHITECTURE.en.md.

Core capabilities

Evidence-gated scene detection

Five meeting types are distinguished by the roles, business actions, and deliverables that appear in the transcript:

• Construction site review — multiple accountable parties (contractor, supervisor, owner) plus a production-control agenda plus periodic review. The formal template is used only when that evidence chain is present.
• General minutes — management reports, business reviews, project weeklies. No fixed template; the model organizes the clearest structure for the source material.
• Personal assistant — personal notes, retrospectives, follow-ups.
• Interview record — candidate evaluation, competency assessment, hiring recommendation.
• Customer visit — customer needs, discussion points, ownership.

The rule-based gate runs first, and it wins: the model cannot classify a meeting as a construction review on the strength of generic words like "project", "schedule", "quality".

Long recordings

• Duration gate with decompression-bomb protection, measured by decoded duration rather than file size.
• Long transcripts are chunked and extracted in parallel; gaps are recorded rather than silently discarding the whole run.
• Job queue with per-user quotas and retries. A dropped connection does not lose work already transcribed.

Export and delivery

• Six templates: construction, general, personal, interview, customer visit, and custom uploads.
• DOCX and PDF export preserving the original layout, with collision-proof filenames.
• Optional DingTalk bot delivery with request signing. Disabled unless configured.

Privacy and security

• No cloud model calls, no API keys, no telemetry.
• Audio, transcripts, minutes, and human edits are stored only on the local machine.
• Login rate limiting, CAPTCHA, signed sessions, CSP and security headers, audit log.
• Template upload hardening: zip bombs, path traversal, and XML entity attacks.
• CI enforces the open-edition boundary: the build fails if cloud-model references or credentials appear in the repository.

Testing

python -m pytest tests-v2                    # 961 tests
python scripts/smoke_test.py                 # offline smoke: scene routing + every export
python scripts/check_open_source_hygiene.py  # open-edition boundary scan

GitHub Actions runs the boundary scan, the backend tests, and the frontend build on every push.

FAQ

Do I need a GPU? No. The FunASR Chinese model and a quantized 7B model run on a 16 GB CPU machine — just slower. An NVIDIA or Apple Silicon GPU speeds things up noticeably.

What if the model download fails? Only the first download needs network access; everything afterwards is offline. Point JKINCO_ASR_MODEL_DIR at an existing model directory to skip the download entirely.

Can I look around before downloading models? Set JKINCO_DEMO_DATA=1 and start. Two example minutes (a construction review and a customer visit) appear in history, and export, search, and templates all work without any model.

Is there live captioning? Experimental. Set JKINCO_REALTIME_LOCAL_ASR=1 to run live captions through a local paraformer-zh-streaming model (an extra download on first use). Off by default.

Can I use it commercially? Yes, under the MIT License. Please keep the copyright notice.

Does it work with languages other than Chinese? The ASR model and the domain lexicon are tuned for Chinese meetings. The architecture is model-agnostic, so swapping in another FunASR or Whisper model is possible — multi-language support is on the roadmap rather than finished.

Roadmap

• Upload → local transcription → scene detection → minutes → export
• History knowledge base and local meeting Q&A
• Streaming live captions (experimental: JKINCO_REALTIME_LOCAL_ASR=1)
• Speaker diarization and role attribution
• WebDAV / cloud-drive backup
• Desktop installers (macOS / Windows)
• Multi-language support (Cantonese / English / Japanese)

Full plan: docs/ROADMAP.en.md.

Contributing

Issues, pull requests, and localization help are all welcome. Please read CONTRIBUTING.md first, and run the boundary scan and the test suite before submitting.

SUPPORT.md states what the maintainer commits to — first response within five working days, security acknowledgement within 72 hours, pull requests reviewed within seven — and, just as importantly, what is out of scope.

License

MIT © 2026 JKinco

---

---

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

Support this project

If JKinco Listen saved you from typing up one more set of meeting minutes, please Star it ⭐.