Scribe

A small Linux/CUDA utility for turning audio files into verbatim Markdown transcripts plus timestamped JSON sidecars. Built around whisper.cpp and ffmpeg. Part of the raibid-labs ecosystem.

What it does

Take an audio file → run it through ffmpeg (resample to 16 kHz mono PCM) → run it through whisper.cpp (CUDA-accelerated on NVIDIA GPUs) → write two files to a configurable output directory:

• <slug>-<timestamp>.md — verbatim transcript wrapped in YAML frontmatter
• <slug>-<timestamp>.json — segment-level timestamps for downstream tooling

Design contract

• Verbatim only. No summarization, no paraphrase, no condensation, no LLM in the path. If you wanted Snipd, this isn't it.
• Single-purpose. Audio in, faithful text out. Annotation, link insertion, search, summarization — those are downstream concerns the user (or another tool) handles.
• Two-file output. Markdown for humans, JSON for tooling. Same base name, same directory.
• Configurable. Output directory, model choice, and whisper.cpp flags surface via CLI flags and (eventually) .fsx config.

See docs/01-architecture.md for the full design rationale.

Status

Phase 0 — repository init. No implementation yet. See issues for the work plan.

Quick start

(Forthcoming once the implementation lands.)

scribe transcribe --input podcast.mp3 --out-dir ~/transcripts

Documentation

• docs/00-index.md — index and overview
• docs/01-architecture.md — design rationale, the verbatim contract, dual-output design, why no LLM is in the path
• docs/02-roadmap.md — planned input adapters, output formats, deferrals, explicit non-goals
• docs/03-related-tools.md — how Scribe fits with other raibid-labs projects (Scryforge, voice-stuff, Phage, gudpkm n8n)
• docs/04-cli-reference.md — subcommands, flags, env vars, exit codes, output layout
• docs/05-contributing.md — dev environment setup, code style, PR conventions, how to extend inputs and outputs
• docs/06-troubleshooting.md — common failure modes and fixes
• docs/07-self-hosted-ci.md — self-hosted GitHub Actions runner for the end-to-end integration test, security model, host hardening

Related raibid-labs projects

• Scryforge — Fusabi-powered TUI information rolodex (RSS, email, YouTube, Spotify, Reddit, bookmarks). Will eventually invoke Scribe as the engine behind a transcribe-to-vault action on podcast and YouTube items.
• voice-stuff (legacy, local) — Python push-to-talk dictation prototype. Shares the underlying whisper.cpp build with Scribe. Slated for a Rust rewrite as Murmur (dictation) plus a separate voice-agent project.
• Phage — context composition engine. Pattern reference for Scribe's eventual .fsx config layer (Fusabi-as-config-DSL).
• gudpkm n8n stack — the voice_memo workflow has a documented TODO to call Scribe for the audio-upload branch of its webhook.

See docs/03-related-tools.md for the full picture.

License

Dual-licensed under either of:

• Apache License, Version 2.0 (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
• MIT License (LICENSE-MIT or http://opensource.org/licenses/MIT)

at your option.