Scriber

What it is. Scriber is a local-first macOS app for meeting transcription and minutes. It records a meeting, transcribes it on-device in real time with word-level timestamps, separates speakers, and walks you from a raw recording to clean, approved minutes — no audio leaves the machine unless you explicitly opt into GPU offload on your own hardware.

Why I built it. Taking good minutes by hand is tedious, and every cloud transcription service means handing confidential meeting audio to a third party. Scriber keeps the entire pipeline — capture, transcription, diarization, and drafting — on the device, so a private meeting stays private by default.

How it works. A two-process design: an Electrobun/React desktop shell driving a Python FastAPI sidecar. Live capture runs MLX Whisper large-v3-turbo on Apple Silicon with streaming partials that refine into final segments. Audio preservation is the prime directive — each clip streams straight to disk independent of transcription, so a crashed inference engine never takes the recording with it. Post-processing is a pipeline: lossless re-transcription → windowed sherpa-onnx speaker diarization (crash-isolated subprocesses with cross-window speaker stitching) → optional per-chunk Gemma refinement → an LLM "Compose Minutes" wizard that consolidates the diarized transcript, your section markers, the agenda, and the org roster into finished minutes, then asks you to resolve attendance, names, and misheard passages before finishing. Every destructive edit snapshots a restorable revision first. An optional Dockerized GPU worker (faster-whisper large-v3 + pyannote 3.1) offloads the heavy pass to a RunPod-style coordinator, with automatic fallback to the fully local path when it is unreachable.

Tech: Electrobun, React/TypeScript/Vite, Python 3.11 FastAPI, MLX Whisper, sherpa-onnx, pyannote, OpenAI-compatible LLM endpoints, Docker/CUDA GPU worker.

Links: github.com/yolanother/scriber

Built Jun–Jul 2026.