Snapshot of the experiment taken October 19, 2026 at 11:27 PM. Numbers update when the page is republished.
CymaticBridge
The future of robot hearing
Bio-inspired cochlea model • Trained listening network • Fold-by-fold validation
EAR LABStage 1
Connecting
Model scores
Clean-audio accuracy, mean of 5 folds x 2 seeds
Experiment progress
0%Overall completion
Each finished training run, in order
Cochlear
transductionCARFAC turns sound into the firing pattern of the inner ear
transductionCARFAC turns sound into the firing pattern of the inner ear
Architecture overview
From sound to answer, the path every clip takes
Sound inESC-10 clips, levelled, at –
Cochlea or melCARFAC ear model, or a mel spectrogram (control)
Ear features10 ms frames, 5 ms frames, or + pitch (SAI)
Listening networkSame small network for every ear; 10 sound classes
Noise test
Accuracy as pink noise gets louder (clean → 0 dB)
Appears when the noise test finishes.
Fold-by-fold validation
Test accuracy per fold, average of 2 seeds; best in each row glows
Noise test
Trained on clean audio, tested on fold 5 with pink noise; in brackets the change from that ear's clean score
One dog bark, both ears
Latest log
Cybernetic learning pipeline // Stage 2
Two listening models trained without labels on ~110 h of sound and speech: one hears through the cochlea model, one through mel. Then the test. Every number here is live from the PC.
Ear pass
Every clip turned into ear pictures, both ears
Training clips finished over time (one point a minute)
Machine
Live load, refreshed every minute
Training
Masked-autoencoder pre-training, same model, steps and seed for both ears; lower loss means it rebuilds hidden sound better
Run 1 · the test
Run 1 (17 to 18 October). Frozen encoders, one linear layer on top: ESC-50, ESC-10 and pink noise; untrained copies show what training added
Run 1 → Run 2: what we learned, what runs next
Every Stage 2 training run and test is logged here. Run 1 result, the quick screens that explain it, and Run 2 (training overnight, 18 to 19 October)
Run 1 verdict
Mel won: 82.7% against 69.9% on ESC-50, and it stayed ahead in noise. Both models learned (untrained copies score 26 to 33%), so the training worked; the cochlea input was the weak link.
Why the cochlea lost
Quick screens scored eleven ways of reading the ear in about 15 minutes. Run 1's cochlea input squashed the nerve signal with a log curve that buries quiet detail. With the ear's own cube-root compression the same signal already ties mel, and beats it in noise.
Closer to the real nerve
The real auditory nerve has three kinds of fibre: very sensitive, medium and hard to excite. CARFAC can model all three. Together with a first "brainstem" step (how fast each band's loudness flickers), this was the best input of all.
| Quick screen (simple scorer, not a trained model) | ESC-50 | Pink 20 dB | Pink 10 dB | Pink 0 dB |
|---|---|---|---|---|
| Mel, as in Run 1 | 53.7% | 37.8% | 25.5% | 6.2% |
| Cochlea, as in Run 1 | 50.1% | 36.8% | 24.0% | 9.8% |
| Nerve, cube-root compression | 54.5% | 44.8% | 32.0% | 7.2% |
| Three fibre types | 56.9% | 46.5% | 35.0% | 9.0% |
| Mel + cube-root + brainstem (fair control) | 58.9% | 52.2% | 42.0% | 12.2% |
| Three fibres + brainstem | 62.4% | 56.2% | 39.5% | 11.5% |
Run 2 resultMel won again: 82.1% against 58.2% on ESC-50 (ESC-10: 93.3% against 84.0%). Both started level untrained (about 36%), but the fibre model's training loss began near 47 against mel's 1: short clips were padded with a "silence" value the sensitive fibres never produce, and the model spent its training rebuilding that padding.
Run 3 resultThe padding fix worked: the three-fibre ear jumped from 58.2% to 76.6% on ESC-50 (ESC-10: 92.2%), beating Run 1's cochlea (69.9%) and closing most of the gap. Mel still leads: 82.6% (ESC-10 93.5%), and it stays ahead in noise (pink 0 dB: 29.5% against 19.5%). The gap went from 24 points to 6.
Run 4, running nowFive quick screens on 19 October: the ear's gain control at half and double strength, 4x faster and 4x slower, and the nerve's fine timing. The gain control changed little; fine timing lifted the fibre input from 56.9% to 60.9% (clean: 51.2% to 58.5%), so Run 4 adds it. The timing pass over the ~110 hours of training audio takes about 6 hours, then training about 3; results expected late on 19 October.
Neural map
What the ear sends the model, one 5 s clip
time → 5 s