Cadence is an Electron desktop application for Windows that records meetings, transcribes them locally using NVIDIA's Parakeet speech-to-text model, optionally separates speakers using pyannote diarization, and produces AI-generated meeting notes. It also functions as a voice dictation tool with global hotkeys and a read-aloud feature using Microsoft Edge neural voices. The frontend is built with React and TypeScript, while the heavy ML work runs in Python worker processes (via FastAPI/uvicorn) that the main Electron process spawns and manages.

It targets a real privacy problem: most meeting recording tools (Otter, Fireflies, Zoom's AI features) upload audio to cloud servers for transcription. This concerns anyone discussing sensitive business, legal, medical, or personal matters. Cadence keeps recording, transcription, and speaker separation entirely on the user's machine. Cloud services (DeepSeek/OpenRouter for note summarization, Edge TTS for read-aloud) are opt-in and clearly flagged as such.

The intended users are Windows-based knowledge workers, consultants, researchers, journalists, and privacy-conscious professionals who need meeting records but don't want their audio sitting on a third party's servers. Users install the app, it runs in the system tray, and captures both microphone and system loopback audio as separate FLAC tracks. On first transcription, it auto-provisions a Python environment via the uv package manager and downloads models — no manual setup required.

The architecture cleanly separates concerns: the Electron main process (src/main/) handles recording orchestration, IPC, settings, hotkeys, and ML worker lifecycle; the preload script exposes a typed API bridge; the React UI (src/ui/) renders meetings, settings, and a transparent dictation overlay window. Audio capture uses electron-audio-loopback for WASAPI system audio without needing a virtual cable. Global hotkeys are handled by uiohook-napi. Python workers in ml/ do actual ASR and diarization via nemo_toolkit and pyannote.audio, exposed via a local FastAPI HTTP interface.

Code quality signals are healthy: strict TypeScript with separate node/web tsconfigs, vitest unit tests present (hotkey-match.test.ts, transcript.test.ts, datetime.test.ts, i18n.test.ts, tts-voices.test.ts), a CI workflow, a CSP layer (security.ts), MIT license, CHANGELOG, CONTRIBUTING guide, and 14 localized UI languages. Nothing in the reviewed code suggests hidden telemetry, exfiltration, or obfuscation — the network calls are exactly the ones the README describes (Hugging Face model downloads, optional LLM APIs, optional Edge TTS).