Vafthrudnir
A C++ companion that reads your own files and measures how each one sounds — rhythm, key, spectrum, structure, dynamics — then hands the coordinates back to the deck. Pure DSP and no machine learning anywhere in it, so it runs on the kind of hardware Tapedeck itself runs on. This is what replaces the Every Noise placeholder behind the Genre Map and Playlist Lab. There is an architecture and there is no code: the repository is empty and nothing is installable yet.
Vafþrúðnir is the giant Odin travels to see in the Vafþrúðnismál to find
out what he does not already know. Spelled Vafthrudnir everywhere it has to be
typed.
The Genre Map and Playlist Lab today are built on Every Noise at Once, which places artists. That is a real map and a useful one, but it means every track by one artist shares the same six sound coordinates — the four listening axes are all that separate them. Tapedeck says so rather than dressing it up as per-recording audio analysis.
It is a placeholder. This is the replacement.
The state of it, plainly
There is an architecture, argued and written down. There is no code, no release and nothing to install, and the repository is empty. It is listed here because the Genre Map and Playlist Lab both name it as what comes next, and because the manifold view is parked until it exists — not because there is something to try.
Several decisions are still open, and they are the interesting ones: whether the loudness meters come from Crete or from Essentia, whether one worker can serve several decks, and whether the work list is everything you have listened to or the whole library.
What it is
A C++ companion that reads your own audio files and produces per-file, pure-DSP descriptors — no machine learning anywhere in it. That is the whole design constraint rather than a detail: the features come out of signal processing, so it runs on the kind of hardware Tapedeck itself runs on instead of demanding a spare desktop. A deck on a Pi Zero 2 W should not need a second machine with four cores and 8 GB of RAM to find out how its own music sounds.
AudioMuse-AI reaches the same goal from the other direction, with learned audio embeddings and the hardware that implies. It exists today and Tapedeck can already connect to it; this is the route that asks nothing extra of your hardware. They are different trades, not competitors.
What it measures
| Group | Some of what comes out |
|---|---|
| Rhythm | BPM and its confidence, beat count, onset rate, danceability — by detrended fluctuation analysis, not a model |
| Tonal | Key, scale and strength under several profiles, tuning frequency, chord histogram and change rate, dissonance |
| Spectral | MFCCs and their covariance, mel and Bark band statistics, centroid, spread, flux, rolloff, flatness, spectral complexity |
| Structure | Intro and outro silence, fades, and the effective duration against the one the container claims |
| Quality | Clipping runs, DC offset, stereo correlation and width, mono sold as stereo |
| Dynamics | Dynamic range, measured by the same code as Crete so the two agree |
Descriptor names follow AcousticBrainz and Essentia's MusicExtractor wherever
one corresponds, so results stay comparable with that corpus rather than being
a private vocabulary.
And an AcoustID fingerprint, which is the one that pays for itself immediately: 57% of a real listening history carries no recording MBID, and a fingerprint is how those tracks get one. That is a metadata win sitting inside an audio-analysis pass, available because the file is being decoded anyway.
Fake-lossless detection is deliberately not in this list. Spectral-cutoff analysis is AudioCheckr's job — a separate, paused project — and letting this become a second one by accident is a named risk rather than an oversight.
A companion, not part of the binary
It is a separate program, and a separate install. The reasons are practical rather than legal — both Tapedeck and Essentia are AGPL-3.0, so there is no friction in either direction. What argues against compiling it in is the arm cross-compile and the engine tick: a Pi 4B running the household instance has five seconds between ticks, and decoding a record inside that budget is not a thing to attempt in-process.
Two further rules follow from the same caution. Each file is analysed in its own child process, because FFmpeg and Essentia can abort on a damaged file and one bad record must not take down the queue. And decoding is streamed in blocks from the first commit: holding a whole track in memory costs about 1.8 GB for ten minutes of 24/192 stereo, which rules out the Pi, a 40-minute raga, and single-file album images all at once.
How it talks to the deck
Vafthrudnir asks for work; Tapedeck never pushes it. It leases a batch of jobs, heartbeats the long ones, and submits results as JSON over HTTP. One mechanism covers every deployment, survives either side restarting, and keeps the deck's side down to a few indexed queries — which is what matters when the deck is on a Pi.
Results are written to a local SQLite outbox before they are sent and deleted only once the deck has acknowledged them, so an upgrade or an outage loses nothing and re-analyses nothing. Re-running after a crash, a retag or a library move is free, because results are keyed on a hash of the decoded audio rather than the file's bytes or its path.
Your media-server credentials never leave Tapedeck. A job carries the paths
the servers reported — Plex's Part.file, Jellyfin's Path, Navidrome's
equivalent — and Vafthrudnir rewrites them through its own map to find the file
on disk. Where there is genuinely no shared storage, it streams the bytes
through the deck as a last resort.
That map is per media server, because every party sees a different path for the same file: a container on a NAS, a deck elsewhere, and Plex reporting its own view of both. So the deck sends a few real paths, Vafthrudnir searches its library roots and proposes the prefix pair, and an admin confirms it — preview then apply, the same rule the metadata sanitiser follows. A path that maps but cannot be read is reported as a permissions problem rather than a mapping one, because those are different answers and one shrug covering both is the bug this codebase keeps fixing.
What it will not do
Absolute, per-file features only. Normalisation, dimensionality reduction and similarity stay in Tapedeck. Anything computed relative to the library goes stale the moment the library changes, and if Vafthrudnir produced it, every new album would invalidate every stored result.
It analyses files, not recordings. One recording can exist as a FLAC and an MP3, a 24/96 and a CD master; deciding which file represents the recording is the deck's join to make, and keeping it out of here is deliberate.
Where the coordinates come from
| Today | Six axes from Every Noise, through the artist. Four from your own listening — era, familiarity, hour of day, recency. |
| Next | Per-recording features computed from the file itself, via a companion sitting beside your Plex, Jellyfin or OpenSubsonic library. |
When it lands, a track gets its own position rather than inheriting its artist's, and "similar" stops meaning "by someone adjacent".
Tracks with no place on the map are dropped and counted, never parked in the middle. That rule survives the change.