Funkhaus

A local, privacy-first take on NotebookLM, written in Rust. Collect sources, ask questions and get cited answers, generate briefings/FAQs/study guides, and turn the material into a German two-host podcast. Everything runs on one Mac: a local OpenAI-compatible LLM server, local TTS and ffmpeg. Source texts never leave the machine unless you explicitly allow an online backend per notebook.

Also in this workspace: Stimmwerk, a second app on the same audio/voice crates for scripted multi-voice audio (voice library, voice cloning with documented consent, music cues, sync to subtitles, LLM-assisted prose-to-script).

Status: personal project, built for my own use on Apple Silicon (macOS). Documentation is partly German. No guarantees, but issues and ideas are welcome.

What it does

• Sources: md, txt, pdf, srt/vtt, html, docx, URLs, whole folders, YouTube captions via yt-dlp (text only).
• Ask: answers with clickable [n] citations into the source text. Retrieval is BM25 plus German-aware fields; optional embeddings.
• Make: briefing, FAQ, study guide, timeline, saved as notes.
• Podcast: LLM writes a two-host script (JSON-schema constrained), an automatic checker flags digits, abbreviations and numbers that are not in the sources, TTS renders line by line with a content-hash cache (edit one line, only that line re-renders), mix with loudness presets, optional jingle and ducked music bed, chapters, show notes, ID3 tags, spoken AI disclosure.
• Agent friendly: the CLI prints JSON on stdout; a Claude Code skill is included in skills/funkhaus/.

Architecture

crates/core      store, import, chunking + BM25, LLM client orchestration, ask/make, podcast pipeline
crates/audio     script format (tags, [pause]), mix timeline, loudness presets, MP3/ID3, chapters, music placement
crates/voice     TTS invocation, content-hash cache, pronunciation lexicon, personas, GPU lock, music generation call
crates/llm       backends (OpenAI-compatible local servers, optional online), keychain-based API keys
crates/cli       `funkhaus` binary
crates/gui       `funkhaus-gui` (eframe/egui)
crates/stimmwerk-{core,cli,gui}   Stimmwerk app and CLI

External tools it calls (not included): a local OpenAI-compatible LLM server (default http://127.0.0.1:8095/v1), an omnivoice TTS CLI, ffmpeg, optionally yt-dlp and an ACE-Step CLI for music.

Measured numbers (from the repo's own docs)

• Loudness presets, measured with ffmpeg ebur128: podcast -16 LUFS / -1.5 dBTP, hoerbuch -19 / -3, video -14 / -1; all within +/-0.5 LU of target, true peak within spec, with and without jingle.
• Retrieval eval (44 questions on two fictional fixture sets, bootstrap CIs): adding German map fields to BM25 raised recall@10 on the solarboot set from 0.364 to 0.909 and MRR from 0.273 to 0.675; on the tiny kleinholm set the difference is within noise.

Quick start

cargo build --release --workspace
cargo test --workspace
./target/release/funkhaus doctor        # checks LLM server, TTS, ffmpeg

funkhaus nb new "Kleinholm"
funkhaus src add kleinholm fixtures/haiku/kleinholm.md fixtures/haiku/chronik.txt
funkhaus ask kleinholm "Wie viele Einwohner hat Kleinholm?"
funkhaus make kleinholm briefing
funkhaus podcast kleinholm --len kurz

scripts/build-app.sh additionally builds a macOS app bundle. Data lives in ~/Library/Application Support/funkhaus/ (override with FUNKHAUS_HOME). The full German manual is HANDBUCH.md.

Requires Rust >= 1.85 (edition 2024).

Fixtures

All fixtures are fictional or synthetic (the village "Kleinholm", sailing transcript "Solarboot", generated seminar texts).

Credits and licenses

Built on open-source Rust crates (see Cargo.lock). Models, TTS voices and music models you plug in come with their own licenses, which apply to their output.

License

MIT, see LICENSE.