bass-foundry/docs/rivet/research/psychoacoustics.md

# Psychoacoustics, expectation, and creative direction

## The product target

Treat a heard musical gesture as the unit of composition: its trigger timing,
pitch movement, articulation, timbral trajectory, placement against other parts,
and relationship to what the listener has already heard. MIDI notes describe
only some of those properties. A held note can contain a rising resonance,
several rhythmic pulses, a change in texture, and a long answering tail.

The design proposals below are hypotheses for a creative tool. The cited
experiments establish particular perceptual effects under their study conditions;
they do not establish universal rules for riddim or Devath's intentions.

## Research-grounded mechanisms

### Expectation depends on context

Cheung et al. found an interaction between uncertainty and surprise in ratings of
Western tonal chord sequences. Both an unexpected event in a predictable context
and an expected event in an uncertain context could receive high pleasure ratings.
This supports exploring relationships between recognition and deviation, while
leaving genre, listener experience, and musical context explicit.

[Cheung et al., 2019, original study abstract](https://cms.mus.cam.ac.uk/publications/cheung-uncertainty-surprise/)

Design inference: preserve an identifiable anchor while changing another
property, and offer confirmation after a passage of ambiguity. A novelty score
alone cannot select a meaningful continuation.

### Groove involves a stable frame and rhythmic challenge

Witek et al.'s drum-break study found greatest pleasure and desire to move at
intermediate syncopation levels. This was a result for their stimuli and listeners,
not a genre-independent target number for a sequencer.

[Witek et al., 2014](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0094446)

Design inference: distinguish the pulse-bearing parts from a freely moving
foreground. A surprising bass or lead event can be easier to follow when drums,
phrase length, or a recurring attack still provide orientation. Intentional
extremes remain valid creative choices.

### Timbre contributes its own tension

Farbood and Price found evidence linking greater roughness, inharmonicity, and
spectral flatness to greater judged tension in their sound comparisons. Evidence
for spectral centroid as a tension predictor was ambiguous. Brightness, roughness,
loudness, harmonic function, and emotional meaning should remain separate concepts.

[Farbood and Price, 2017](https://bpb-us-e1.wpmucdn.com/wp.nyu.edu/dist/f/11865/files/2020/08/Farbood_Price_2017.pdf)

Design inference: a melodic arrival and a timbral release can happen together,
independently, or in opposition. A line can keep its pitches while its texture
becomes less abrasive. A quiet sound may remain harmonically unresolved.

### Grouping determines which sounds form a phrase

Elhilali et al.'s human psychophysics experiments showed that relative timing can
strongly affect whether tones are heard as a coherent stream. Their findings
challenge an account based solely on frequency separation.

[Elhilali et al., 2009](https://pmc.ncbi.nlm.nih.gov/articles/PMC2673083/)

Design inference: a motif distributed between sounds needs perceptual connections.
Shared timing, related articulation, recurring contour, and deliberate handoffs
are candidates to test. A MIDI line split across unrelated patches may be heard
as several parts rather than one melody.

### Preferences are learned as well as sensory

McDermott et al. found substantial cultural variation in consonance preferences,
including indifference to consonance versus dissonance in a population with little
Western musical exposure. This argues against presenting Western consonance
preferences as a universal measure of musical worth.

[McDermott et al., 2016](https://www.nature.com/articles/nature18635)

Design inference: genre fluency and the producer's taste matter. Scale adherence,
chord-tone density, and estimated roughness can describe an option; they should
not become an automatic good/bad ranking.

## Reading the supplied Devath references responsibly

The [local measurements](devath.md) establish changes in mix
level, attack activity, and broad spectral texture. For example, HENKA has a
measured reduction around 2:03–2:35 and Phantasy around 1:27–2:23. Petals maintains
more consistent relative activity across much of its earlier duration. These
contrasts support studying both large departures/returns and development within
an ongoing groove.

Those measurements do not establish what the melody is, how the harmony moves,
which sound is foreground, or what emotional meaning a listener assigns. Exact
note-level claims require an annotated listening pass, preferably with stems or
MIDI. The existing report deliberately leaves those questions open.

For each reference passage, annotate the following rather than only naming a key:

1. **Anchor:** what remains recognizable—attack pattern, interval, texture, bass
   pulse, or a small melodic shape?
2. **Foreground:** which part attracts attention, and when does that change?
3. **Motion:** does movement come from pitch, amplitude pulses, spectral movement,
   articulation, spatial effects, or several at once?
4. **Expectation:** what seems about to happen, and what is confirmed or withheld?
5. **Space:** which event is absent, shorter, farther away, or easier to hear?
6. **Memory:** when something returns, what has changed around it?

These are listening questions, not findings already established about each file.
Reference tracks supply examples of possibilities; they do not define the bounds
of the producer's own style.

## What the tool should let a producer do

### Work on relationships

Useful requests are specific and relational:

- Keep this bass groove and let a smaller melodic answer occupy its gaps.
- Preserve the pickup; make the destination later and more exposed.
- Keep the notes but contrast short, dry articulations with a sustained answer.
- Let one timbre finish a phrase that another timbre starts.
- Keep harmonic ambiguity while reducing rhythmic activity.
- Return to a familiar gesture with softer texture and different accompaniment.
- Turn the ending into a pickup for the next loop, rather than closing it.

The original motif can survive any of these transformations. The producer chooses
which relationship matters and which dimensions are free.

### Separate useful descriptors

Avoid a single emotional-intensity control. Explore these dimensions separately:

| Dimension | Example experiment |
|---|---|
| Pulse clarity | Keep drums fixed while moving foreground attacks. |
| Anticipation | Bring the pickup forward; delay the destination. |
| Activity | Remove connective notes while retaining signature anchors. |
| Articulation | Keep onset locations and compare different note lengths/tails. |
| Harmonic stability | Hold melody fixed over different bass/chord contexts. |
| Timbral tension | Keep pitches fixed while changing an explicitly mapped patch macro. |
| Continuity | Pass the motif between related and contrasting voices. |
| Familiarity | Reintroduce a recognizable fragment after a departure. |

These descriptors help propose experiments. They do not measure longing, sadness,
euphoria, or artistic merit. Several can point in different directions at once.

### Capture how the actual sound behaves

The same MIDI onset can produce different heard attacks through different patches.
A slower attack, rhythmic modulation, portamento, or a long release can change the
musical role of a note. Accurate evaluation therefore needs the production sound
and surrounding loop, with repeated audition rather than only an isolated tone.

A future sound profile could record an approximate attack delay, release tail,
comfortable register, and user-mapped expressive controls. Audio returned from the
synth could help measure those responses. A MIDI-only generator cannot infer the
behavior of an arbitrary patch or automatically control its macros without an
explicit routing/mapping arrangement.

### Make creative moves composable

The current eight families are useful initial experiments. Further freedom comes
from chaining smaller operations with a visible history:

- select and preserve an anchor;
- fragment, displace, stretch, repeat, or subtract an event group;
- choose a destination relative to the surrounding harmony;
- hand a gesture to another voice;
- change its articulation or a mapped timbral trajectory;
- bring a remembered fragment back in a new context.

A producer should be able to branch an idea, pin a particular decision, compare
branches in context, and reuse a personal gesture. Multiple operations should
compose without forcing every phrase into a statement/climax/resolution arc.
Circular, suspended, sparse, and continuously evolving structures are valid.

### Preserve learning and authorship

Three proposals should illustrate different musical hypotheses. For example:

1. same pitches, more space, a later arrival;
2. same groove, a different harmonic destination;
3. a fragment handed between two explicitly assigned voices.

Keep the surrounding loop and listening level comparable, so the producer can
judge the changed relationship. Store the chosen material and the specific
choices he wants to preserve. Optional explicit favorites can become a personal
vocabulary; generic musical rules should remain adjustable.

## Practical next research steps

1. **Annotated reference sessions.** Select short passages from the supplied files
   and mark anchors, foreground roles, handoffs, withheld events, and returns.
   Check claims against repeated listening instead of treating a waveform as a
   melodic transcription.
2. **Controlled comparisons using the friend's own patches.** Keep notes fixed
   while changing articulation; keep articulation fixed while changing rhythm;
   compare the same line under two harmonic settings. Check the whole loop as
   well as isolated examples. Avoid making the louder example automatically seem
   more successful.
3. **A reusable gesture and branch history.** Extend the current note/pin model
   toward saved groups with independent timing, pitch, articulation, voice, and
   mapped modulation choices.
4. **Two-role experiments.** Explore interlocking bass and answer lines before
   expanding into a full arranger or arbitrary multi-instrument generator.
5. **Producer evaluation.** Ask whether a suggestion leads to a usable section,
   whether he can retain its useful moment, and whether the tool keeps him in the
   writing flow. Test over repeated sessions; mathematical novelty is insufficient.

## Current implementation boundary

Version 0.2 already covers structured note/rhythm proposals, harmony context,
separate note locks, pinned sections, and external MIDI audition. It does not yet
represent patch-dependent audible gestures, mapped timbral motion, multiple sound
roles, reusable branch histories, or verified perceptual/emotional judgments.
The next design work should make those relationships explicit while keeping the
producer's listening decisions central.