natural-perspective-spatial-audio 0.1.2

Turn any recording into a 7.1 surround mix: a model designs a per-song spatial placement and renders a lossless 8-channel FLAC.

Natural Perspective Spatial Audio

Turn any recording into a 7.1 surround mix. After a song is split into its instrument stems, a model invents a scene and decides where each instrument sits around you — then a deterministic renderer builds a lossless 8-channel (7.1) FLAC for your media server.

One real mix: every stem placed on the stage, the crowd behind you.

See it — no install

Open examples/index.html in any browser, on any OS. It's a finished mix's scene, soundstage, routing, and stem levels — the actual output of the tool.

Run it

Easiest — install from PyPI with pipx, no clone needed. Needs Python 3.10+ and a system FFmpeg:

brew install ffmpeg                   # macOS; Linux: sudo apt install ffmpeg
pipx install 'natural-perspective-spatial-audio[full]'
spatial-standards-gui                # or:  spatial-standards song.flac

Or from a clone — one command (macOS/Linux) sets up a private virtualenv and installs everything; on a Mac with Homebrew it installs Python for you too:

./install.sh
./gui                                # the GUI   (or: .venv/bin/spatial-standards song.flac)

Manual / Windows

Needs Python 3.10+ (3.12 recommended — widest wheel coverage):

python3 -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install '.[full]'                # quote it — the [..] is a shell glob otherwise
spatial-standards song.flac          # also: a folder, or a URL
spatial-standards-gui                # or the GUI

The GUI uses Tkinter: it ships with the python.org installer (recommended on macOS), with Homebrew add python-tk, on Debian/Ubuntu apt install python3-tk.

[full] brings everything as Python packages — FFmpeg + ffprobe (via static-ffmpeg), Demucs, the crowd model (audio-separator), and yt-dlp — so a fresh machine works after one install, no system setup. It's CPU by default; for an NVIDIA GPU install a CUDA build of PyTorch from pytorch.org and pip install 'audio-separator[gpu]' (much faster). The first run downloads model weights (a few hundred MB).

Prefer your own tools? pip install . has no required Python packages and just calls ffmpeg, demucs, audio-separator, and yt-dlp from your PATH.

The model layer needs an ANTHROPIC_API_KEY; without one the tool falls back to a built-in mix and runs fully offline.

Output drops straight into Plex/Jellyfin/Kodi:

<Artist>/Natural Perspective Spatial Audio/<Title> [...].flac   (+ per-album index.html)

How it works

Separate stems → measure each stem's level → a model invents a scene and emits a full mix configuration → a safety-clamped renderer builds the 7.1 FLAC. No audio is uploaded — only metadata, cover art, and the measured stem levels inform the design. Every track saves its config and an index.html documenting the scene, routing, and the exact model prompt/response.

Your responsibility, and credits

By processing a file or URL you affirm you have the right to it. This tool hosts no content and ships no audio. See NOTICE for that and full credit to the projects it builds on — FFmpeg, Demucs, audio-separator, yt-dlp, and the Mel-Band RoFormer crowd model.

License

Apache-2.0 — see LICENSE and NOTICE. Provided as is, without warranty of any kind.