闻道 · Wendao

Self-hosted meeting & media transcription with speaker diarization and AI minutes.

A privacy-first, fully self-hosted alternative to Feishu Minutes / Otter.ai — your recordings, your GPU, your data.

English · 简体中文

Note

Wendao runs entirely on your own infrastructure. Transcription uses open Whisper-family models on your GPU; AI minutes use your local Ollama. No audio, transcript, or meeting content ever leaves your servers.

Transcript with word-level karaoke highlighting, speaker diarization, a synced player, and on-demand AI minutes — all self-hosted.

Library — folders, full-text search, and processing status at a glance.

✨ Features

• 🎙️ Upload → transcribe. Drop in audio or video; get an automatic transcript with word-level timestamps.
• 🗣️ Speaker diarization. Cluster speakers by voiceprint, rename Speaker 1 → real name, and re-assign segments.
• ▶️ Synced playback. Click any word to jump the player; karaoke-style word highlighting; playback speed, skip-silence, captions (CC) and bilingual captions.
• 🤖 AI minutes (local LLM). On-demand summary + key points + risks, chapter breakdown, action-item extraction (with owners), and auto-title — each with clickable source citations back to the transcript.
• 💬 Chat with your transcript. Ask questions and get answers grounded in the recording, with cited line numbers.
• 🤝 Collaboration. Highlights, per-segment comments, shareable clips, and link-scope permissions.
• ✍️ Editing. Fix transcript text and tidy up paragraphs directly.
• 🌐 Translation. Translate transcripts (zh / en / ja) and produce bilingual transcripts.
• 📊 Organize & analyze. Folders, full-text search, in-transcript search; per-speaker talk-time / share / word-count and access/comment stats.
• 📤 Export. TXT / SRT / Markdown / DOCX (speaker labels and timestamps optional).

AI minutes are generated on demand (not as a blocking pipeline step), so the transcript is available the moment segmenting finishes — the LLM never delays it.

🏗️ Architecture

                          ┌──────────────────────────────────────────┐
                          │  Browser (Next.js 16 web · :3101)          │
                          └───────────────┬──────────────────────────┘
                       same-origin proxy  │  (large uploads go direct)
                          ┌───────────────▼──────────────────────────┐
                          │  API · Fastify + tRPC v11 (:3100)          │
                          │  + pipeline worker                         │
                          └───┬─────────┬──────────┬──────────────────┘
                   Postgres   │   ASR   │   Ollama │   local storage
                  (16 tables) │  :9400  │  :11434  │  (media + artifacts)
                              ▼
                   Python FastAPI · WhisperX / faster-whisper + diarization

Five workspace packages (pnpm + Turborepo):

Package	Role	Port
apps/miaoji	Next.js 16 / React 19 web frontend	3101
packages/miaoji-api	Fastify 5 + tRPC v11 + Drizzle backend (+ pipeline worker)	3100
apps/miaoji-asr	Python FastAPI ASR + diarization microservice	9400
packages/miaoji-contracts	Shared Zod schemas — the cross-language source of truth	—
packages/miaoji-web-ui	Shared React UI components	—

The directory/package codename is miaoji; the product is Wendao. They map 1:1.

See docs/architecture.md for the full design.

🚀 Quick start

Prerequisites

• Node ≥ 20 and pnpm ≥ 9 (corepack enable)
• PostgreSQL 16 (Docker Compose included)
• ffmpeg / ffprobe on PATH
• Ollama with a model pulled (default qwen3:30b-a3b) — for AI minutes / translation / Q&A
• For transcription: Python 3.11/3.12 and a CUDA GPU (CPU works but is slow). See apps/miaoji-asr/README.md.

Run with Docker Compose

git clone https://github.com/wuji-labs/wendao.git
cd wendao
# Create your .env from .env.example and review the defaults.

# Node stack (postgres + api + worker + web)
docker compose up -d postgres api worker web
docker compose exec api pnpm -F @wuji/miaoji-api db:migrate

# The GPU ASR service is optional and profile-gated (needs the NVIDIA runtime):
# docker compose --profile asr up -d asr

Open http://localhost:3101.

Run manually (dev)

pnpm install
# Create your .env from .env.example (see docs/configuration.md).

docker compose up -d postgres                 # or your own Postgres
pnpm -F @wuji/miaoji-api db:migrate
pnpm -F @wuji/miaoji-api db:seed              # dev user
pnpm db:seed:demo                            # optional: a ready-made demo minute

pnpm -F @wuji/miaoji-api dev                  # API     :3100
pnpm -F @wuji/miaoji-api dev:worker           # worker
pnpm -F @wuji/miaoji-web dev                  # web     :3101

# ASR service (separate process; needs GPU + models)
cd apps/miaoji-asr && ./run.ps1               # or: uvicorn app.main:app --host 0.0.0.0 --port 9400

Full self-hosting guide: docs/self-hosting.md · Configuration reference: docs/configuration.md.

🔄 How it works

Each upload runs through a state machine:

UPLOADING → TRANSCODING (ffmpeg) → TRANSCRIBING (ASR) → DIARIZING → SEGMENTING → READY

AI minutes (summary / chapters / to-dos) are then generated on demand via the local LLM. If the LLM is unavailable, the transcript still works — AI features degrade gracefully. The same is true for diarization: if no diarization backend is configured, you still get a full transcript without speaker labels.

🧩 Tech stack

Next.js 16 · React 19 · Fastify 5 · tRPC v11 · Drizzle ORM · PostgreSQL 16 · Zod · Turborepo · FastAPI · WhisperX / faster-whisper · pyannote / DiariZen · Ollama.

📁 Project structure

wendao/
├─ apps/
│  ├─ miaoji            # Next.js web frontend
│  └─ miaoji-asr        # Python ASR + diarization service
├─ packages/
│  ├─ miaoji-api        # Fastify + tRPC + Drizzle backend (+ worker)
│  ├─ miaoji-contracts  # Shared Zod schemas (SSOT)
│  └─ miaoji-web-ui     # Shared React UI
├─ docs/                # architecture, self-hosting, configuration, models, roadmap
└─ docker-compose.yml

⚠️ Status & limitations

Wendao is pre-1.0 and was built for self-hosted/trusted-network use. Please read before deploying publicly:

• Authentication is intentionally minimal — a simple x-user-id header suitable for a trusted LAN. Put a real auth layer / reverse proxy in front before exposing it to the internet.
• The UI is currently zh-CN only. English/i18n is a welcomed, high-value contribution — see docs/roadmap.md.
• Code comments are a mix of Chinese and English (the project originated in a Chinese-speaking team).
• ASR models download on first use and need a CUDA GPU for practical speed.

See docs/roadmap.md for what's planned and where help is wanted.

🤝 Contributing

Contributions of every size are welcome — see CONTRIBUTING.md. Please also read our Code of Conduct. For security issues, follow SECURITY.md (do not open public issues).

📜 License

MIT © 2026 WUJI (wuji-labs).

Model licenses are separate. The default diarization path (DiariZen meeting-base) is MIT, but some optional DiariZen weights are CC-BY-NC-4.0 (non-commercial) and pyannote requires accepting its terms with a Hugging Face token. See docs/diarization-and-models.md before commercial use.

🙏 Acknowledgements

Built on the shoulders of OpenAI Whisper, WhisperX, faster-whisper, pyannote.audio, DiariZen, and Ollama. Product inspiration from Feishu Minutes and Otter.ai.

Made with care by WUJI Labs.