Skip to content

HEARING.md

Latest commit

 

History

History
70 lines (52 loc) · 3.02 KB

HEARING.md

File metadata and controls

70 lines (52 loc) · 3.02 KB
Raw

Unified hearing research direction

This fork is the canonical development home for the combined deepSTRF + libaural research program.

The project combines four complementary sources:

  • upstream deepSTRF for neural datasets, encoding models, recurrence, metrics, checkpoints and training;
  • libaural for auditory obligations, biological teacher roles, adversarial tests and mechanism compression;
  • the Differentiable Auditory Loop (DAL) for representation-guided inverse design;
  • torch_amt for differentiable PyTorch auditory and psychoacoustic mechanisms.

Source state at consolidation

  • deepSTRF base: urancon/deepSTRF@5dba99f60a5cb51548f9f9769941cca377771d01
  • libaural source: dissonance-git/libaural@c280ea60a54c23f9b64a1d8af729e2c0d428a266
  • torch_amt donor: StefanoGiacomelli/torch_amt@aa84b5e9b2e065c06ce196c89306a30c324d80a0
  • DAL paper: arXiv:2606.04103

The DAL paper's published code path is currently unavailable, so no DAL source code is copied here yet. Its differentiable comparison method is retained as a research pattern.

One project, different roles

deepSTRF
neural data + learned response models + recurrence
        +
libaural
auditory obligations + provenance + counterexamples + compression
        +
torch_amt
auditory model components
        +
DAL-style inverse design
        ↓
unified artificial-hearing research system

Neural prediction is a teacher, not the definition of hearing. A biological model is a teacher, not an automatic runtime dependency. A representation must survive the obligation being tested before it can be treated as sufficient for that obligation.

First native migration

deepSTRF.hearing.temporal contains the first migrated libaural AUD-REP-001 ruler. It tests whether a time-resolved representation retains exact reversal and opposite-event-order structure while keeping whole-sequence pooling visible as a destructive control.

The intended migration sequence is:

  1. AUD-REP representation and compression tests.
  2. Original ICNet public evaluation and reliability controls.
  3. StateNet and other deepSTRF learned-state challengers under the same obligation surface.
  4. torch_amt peripheral, modulation, loudness and binaural mechanisms.
  5. Differentiable inverse design using any compatible auditory teacher.
  6. Mechanistic central-path teachers when a named experiment requires them.
  7. Persistent auditory organization, memory, prediction and uncertainty.

Transfer rule

rich teacher or measured data
→ fixed auditory obligation
→ perturbation / counterexample
→ identify the transformation that matters
→ compress it
→ test the same obligation again
→ compare state, latency and compute cost

The research repository may remain much richer than the eventual runtime implementation.

Lineage

dissonance-git/libaural remains preserved as research lineage while its useful experiments and laws move here. New canonical implementation work should land in this deepSTRF fork so the four research lines do not grow into parallel competing architectures again.