run record~/research/cliste/runsAll Cliste runs

Type a section name and press Enter. Use “help” for the list.

Figures: softened for dark viewingDisplay only; image files are unchanged.
← All Cliste runs

Cliste experiment

Improve audio robustness while retaining native visual memory

Execution: completedScientific verdict: partial

The question

Compare three acoustic front ends and two VSA mappings; qualify one frozen choice on new recordings.

Implementation and comparison details

Versioned gain/PCEN candidates, recoverable level, batched native recall and development-only confidence calibration.

What happened

Selected relative_logmel-phase; qualification verdict partial. Inspect per-condition improvements, losses and remaining gaps.

Results and visual artifacts

Open a figure for its input, comparison, and result. Click an image to enlarge it.

Qualification by acoustic condition
Qualification by acoustic condition
Same 48 qualification events and 108 stored encounters. Both policies were chosen only on development. Challenge conditions are diagnostics; false acceptances are recorded separately.
qualification_matrix1999-03: gain_half
qualification_matrix1999-03: gain_half
Observer input, expected native readback, original-encoder candidate and selected-encoder candidate. Refused candidates are displayed for inspection, not accepted answers. Examples include improvements and failures when present.
qualification_matrix1999-08: gain_half
qualification_matrix1999-08: gain_half
Observer input, expected native readback, original-encoder candidate and selected-encoder candidate. Refused candidates are displayed for inspection, not accepted answers. Examples include improvements and failures when present.
qualification_neural_network-05: shift80
qualification_neural_network-05: shift80
Observer input, expected native readback, original-encoder candidate and selected-encoder candidate. Refused candidates are displayed for inspection, not accepted answers. Examples include improvements and failures when present.
qualification_neural_network-06: shift80
qualification_neural_network-06: shift80
Observer input, expected native readback, original-encoder candidate and selected-encoder candidate. Refused candidates are displayed for inspection, not accepted answers. Examples include improvements and failures when present.

Limits and next question

Simulated acoustic channels and supplied event windows. Skala is not built; no physical microphone or room qualification. No semantic category labels or neural training.

Next question. Can confidence use temporal evidence without losing changed-cue recall? Larger offsets and strong noise remain open before future hardware work.

Inspect the evidence

Download the methods and measurements for this experiment. Large training and saved-state files are not included.