Your meetings, transcribed on your own machine. Your interviews, answered live.

A local, bot‑free meeting notetaker and live interview copilot for Windows.
Records your mic and your computer audio, transcribes and separates speakers on your GPU,
then writes the notes, or drafts your answer while the question is still being asked.

Quick start · Features · How it works · Models · Privacy · Desktop build · Contributing

✨ Why Murmur

🎛️ Two modes, one app

Switch between them from the header at any time.

🚀 Features

🧠 How it works

flowchart LR
    subgraph IN["🎙️ Capture"]
        MIC[Microphone]
        SYS[System audio<br/>loopback]
        SCR[Screen share]
    end

    subgraph LOCAL["🔒 Runs locally (GPU / CPU)"]
        AEC[DTLN‑aec<br/>echo cancel + denoise]
        VAD[Silero VAD<br/>segmentation]
        ASR[faster‑whisper<br/>or Parakeet]
        DIA[sherpa‑onnx<br/>diarization]
        STORE[(Local storage<br/>data/)]
    end

    subgraph AI["☁️ Your AI provider (text only)"]
        SUM[Summary,<br/>sentiment, fixes]
        COP[Live answers]
    end

    UI[🖥️ Web UI<br/>127.0.0.1:8765]

    MIC --> AEC
    SYS --> AEC
    AEC --> VAD --> ASR --> DIA --> STORE
    SCR --> STORE
    DIA -- transcript --> UI
    STORE -- transcript text --> SUM --> UI
    DIA -- question text --> COP --> UI

    classDef capture fill:#3157d5,stroke:#3157d5,color:#fff
    classDef local fill:#1f2937,stroke:#7490ff,color:#fff
    classDef cloud fill:#7490ff,stroke:#3157d5,color:#fff
    classDef ui fill:#19191c,stroke:#7490ff,color:#fff
    class MIC,SYS,SCR capture
    class AEC,VAD,ASR,DIA,STORE local
    class SUM,COP cloud
    class UI ui

A FastAPI + uvicorn backend on 127.0.0.1:8765 serves a vanilla‑JS frontend and streams the live transcript and answers over a WebSocket. Audio capture, resampling (soxr), VAD, recognition, diarization and echo cancellation all run in‑process. The design decisions, tuning and latency benchmarks are recorded in docs/adr-001-audio-pipeline.md.

🏁 Quick start

✅ Requirements

1 · Get the code

git clone https://github.com/muradiodev/murmur-notetaker-copilot.git
cd murmur-notetaker-copilot

Download a static Windows build of ffmpeg and put ffmpeg.exe into bin/.

2 · Install (one click)

python -m venv .venv
.venv\Scripts\python -m pip install -r requirements.txt
.venv\Scripts\python -m app.setup_models     REM downloads models on first run
.venv\Scripts\python -m app.server           REM serves http://127.0.0.1:8765

3 · Add a key

Open Settings (bottom left) → API keys, paste a key and choose a Summary provider. Settings, keys included, are saved to data/settings.json and survive restarts.

[!TIP] Want to see what a finished Copilot session looks like before your first call? python tools/demo_copilot.py creates four demo sessions with a fictional candidate; --remove deletes them again.

🤖 AI providers

Provider	Used for	Notes
Anthropic Claude	Summaries, Copilot answers	Haiku, Sonnet or Opus
OpenAI	Summaries, Copilot answers	Model list is read from your own account
NVIDIA NIM	Summaries, Copilot answers	Model list is read from your own account

Every provider has a Test button that checks the selected model responds before you rely on it in a call.

📦 Models

All speech models run locally and live in models/ (not tracked in git). app/setup_models.py downloads what's missing on first run.

🗂️ Project structure

murmur-notetaker-copilot/
├── app/                  # backend
│   ├── server.py         #   FastAPI app, REST + WebSocket
│   ├── audio.py          #   mic + loopback capture, resampling
│   ├── transcriber.py    #   faster-whisper / Parakeet engines
│   ├── diarize.py        #   speaker separation (sherpa-onnx)
│   ├── aec.py, dtln.py   #   echo cancellation
│   ├── summary.py        #   AI summaries, sentiment, provider clients
│   ├── copilot.py        #   live interview answers, resume/document library
│   ├── names.py          #   speaker names from self-introductions
│   ├── storage.py        #   meetings on disk, exports
│   └── static/           # frontend: index.html, app.js, copilot.js, style.css, i18n/
├── tools/                # build_playbook.py, demo_copilot.py
├── packaging/            # PyInstaller + Inno Setup desktop build
├── docs/                 # architecture decision records
├── brand/                # logo and brand kit
├── desktop.py            # WebView2 desktop window
├── setup.bat · run.bat
└── models/ · data/       # local models and user data (git-ignored)

🪟 Windows desktop app & installer

A packaged build with its own window (WebView2) and a setup.exe installer:

powershell -ExecutionPolicy Bypass -File packaging\build.ps1

Requirements and caveats are in packaging/README.md.

🔐 Privacy

[!IMPORTANT] data/ and models/ are git‑ignored. Never commit data/settings.json: it holds your API keys.

🤝 Contributing

Issues and pull requests are welcome.

1. Fork the repo and create a branch: git checkout -b feature/my-change
2. Keep changes focused, and match the style of the surrounding code
3. Run the app and check the change in both modes
4. Open a pull request describing what changed and why

UI strings live in app/static/i18n/en.js, so adding a language starts there.

🙏 Built on

faster‑whisper · sherpa‑onnx · pyannote · WeSpeaker · Silero VAD · DTLN‑aec · NVIDIA Parakeet · onnx‑asr · FastAPI · Geist · Phosphor Icons

📄 License

Released under the MIT License. Third‑party models keep their own licenses; check each model's page before redistributing them.

[!NOTE] Copilot is an interview‑preparation and live‑assistance tool. Use it in line with the rules of any interview or assessment you take part in, and let people know when you record a conversation where the law requires it.

_{Made with 🎧 for people who'd rather listen than take notes.}