Sonokinetic

Can motion in a field of sound be perceived without localizing anything in it?

The instrument builds a field of virtual sources from a piece of audio, gives the field a motion, and makes each source individually too diffuse to locate. Whether a listener still hears the field move is the experiment.

Jonas Techmanski · jonas030405@gmail.com

Eight variants are playable on headphones at techmanskijonas.github.io/sonokinetic.

Setup

python -m venv .venv
.venv\Scripts\activate
pip install numpy scipy soundfile matplotlib sounddevice fastapi uvicorn python-multipart pytest

sofar is optional and only needed for measured HRTFs.

Run

python app.py

Open http://127.0.0.1:8000 on headphones. Switch off operating-system spatial audio first, because it applies a second head model to signals that already carry one.

Audio

Recordings stay out of the repository, so the first run writes demo-drone.wav, 45 seconds of synthetic sustained material, and selects it. The signal is built to what the treatment needs. It has no transients, because onsets restore the localization the treatment removes. It carries energy up to 8 kHz, because the level difference between the ears is a high-frequency cue. Running make_demo_audio.py writes that file and demo-strings.wav beside it.

Any track can be dropped onto the waveform or put in the project folder. Sustained material with few transients works best. Pads, drones, bowed strings, organ, held voice and cymbal wash all qualify. The reason is measured and sits under Sustained material decorrelates far more readily in the Research chapter.

The ? beside the title runs a guided tour. There are three, covering an overview, a hands-on build, and blind testing.

Measured HRTFs

The head model is a rigid sphere by default. A sphere renders mirrored azimuths identically and so carries no front-back information. A measured set supplies it, and the files are free to download:

pip install sofar
mkdir hrtf
curl -o hrtf/mit_kemar_normal_pinna.sofa https://sofacoustics.org/data/database/mit/mit_kemar_normal_pinna.sofa

That is the Gardner and Martin KEMAR measurement, 1.2 MB, 72 positions on the horizontal plane. Set hrtf_file on a variant to use it and leave it unset for the sphere. Direction comes from the measured set while distance stays with the geometry, so near-field behaviour is unaffected.

Interpolation is nearest-neighbour on a 5 degree grid, which suits static sources and is still too coarse for moving ones.

How it works

Audio is summed to mono and rendered as a field of virtual sources. Work is organised in three levels:

• A passage is a time region of the track.
• A variant is one treatment of that passage, or the untreated signal.
• A component is a lattice of sources with a motion field applied.

There are two lattices. A polar one is concentric rings. A grid is a rectangle of sources spanning an extent in metres. Rotation, radial flow and translational drift combine freely on either, so a whirlpool is a polar lattice turning while it flows inward, and driving through a field of sources is a grid with a backward drift. Sources wrap within the lattice, which holds the count constant.

Each component carries its own decorrelation or inherits the variant's. Any share of its sources can be given over to random wandering in place of the coherent motion, following the coherence manipulation in random-dot kinematograms.

Distance is set per component in metres and may reach zero. At the head centre both ears are equidistant, the interaural differences vanish, and the source is heard as a centred, in-head image.

Every variant of a passage is rendered over the same span, aligned to the sample and matched in loudness. During playback the track runs untreated while the variants run alongside it in step, and holding a number key brings one forward.

Reading the results

Two findings constrain how the instrument is used. Both were measured here.

A single render carries no evidence that a field is moving. IACC responds to diffuseness as well as to motion, and modulation at the motion rate is confounded with the material's own periodicity. Over a 20 s passage a completely static field scored 2.04 at 0.5 Hz where moving fields scored 2.65 to 3.14. Every motion claim needs a matched control that is identical except for having its motion stopped. Stop it by setting time_scale to zero, which holds each source where it was at the level it had. Zeroing the rates instead changes the level distribution, because the edge fade keys off whether the lattice wraps. The instrument detects those pairs and reports the difference.

The measures have a noise floor larger than several effects under study. IACC at full decorrelation varies with a standard deviation of 0.038 across seeds. Ring geometry moves it by 0.04 to 0.05, which falls inside that. Hold the seed fixed when comparing, or average across several.

Learning it

Courses run from two ears and two differences through to conducting a listening test that counts as evidence. Research covers the question, where it sits in the literature, the results measured here, and what blind listening has shown. Glossary defines every term the interface uses, split into established material and this project's own vocabulary, each entry labelled by status:

Status	Meaning
Established	Standard material with a citation
Implementation	A design choice in this codebase
This project's term	Vocabulary coined here, not a term of art
Measured here	A result reproducible from the test suite
Conjecture	A hypothesis under test

The labels matter when writing for other people, because motion in unison, motion coherence share and component come from this project and not from the literature.

Files

File	Role
ringfield.py	the DSP, and the file to read first
app.py	web backend, wrapping the renderers and holding no DSP of its own
static/	the interface
courses.json, purpose.json, encyclopedia.json	all written content
tests/	311 tests, to run before and after any change
make_track.py, compare.py, variations.py, sweep.py	command line renderers
make_demo_audio.py	synthesises the two demonstration beds
build_demo.py	renders everything the public page serves
docs/	the public page, whose data/ directory is generated

pytest -q

Rebuilding the public page

python make_demo_audio.py
python build_demo.py

The first writes the two beds. The second renders every variant of each and writes the audio, the monitor traces and the measured numbers into docs/data/. Both beds are synthesised, which keeps the published page clear of material that cannot be redistributed.

The page renders the ring at the geometry the blind sessions used, which is nine sources at two metres with allpass decorrelation and every source wandering 60 degrees at a quarter hertz. A plainer ring with fewer sources and no wander is a much weaker stimulus, so the demonstration would understate what the sessions found.

Each variant is rendered unnormalised, loudness-matched to the untreated reference, then scaled together with every other variant by one common factor. A listener moving between two of them hears a difference of treatment and nothing else.

Known limits

The head model is a rigid sphere by default, with no pinna, so it carries neither elevation nor front-back discrimination. Azimuth 0° and 180° produce identical signals and the direction of motion cannot be recovered from them. A measured HRTF in SOFA format supplies both and is selectable per variant, though interpolation between measured directions is still nearest-neighbour and suits static sources better than moving ones. Distance is carried by level and near-field level difference, with no propagation delay, Doppler, air absorption or reverberation.

Blind listening found the sensation of movement surviving a sum to mono, which removes every interaural difference. The effect is therefore substantially spectral, and calling it interaural would overstate what has been shown.

Everything measured so far comes from one listener who holds the hypothesis. The blind test logs to CSV with full parameter provenance, so adding listeners needs no further work.