Skip to content

Latest commit

 

History

History
200 lines (149 loc) · 9.94 KB

File metadata and controls

200 lines (149 loc) · 9.94 KB

Evaluation Protocol

How anything in PRELUDE gets measured. Two tiers: cheap automated metrics that run on every change, and scarce human listening sessions that are the actual ground truth.

The governing constraint: human listening time is finite, fatiguing, and — and where hearing loss is progressive, diminishing. Every automated metric in Tier 1 exists to ration Tier 2, never to replace it.


Tier 0 — Safety (non-negotiable, runs before any human hears anything)

Enforced in code, in the pipeline, with no bypass flag.

  1. Loudness normalization to a documented target (default −23 LUFS integrated) across all stimuli in a comparison. Unmatched levels invalidate subjective comparison — loudness differences dominate every other perceptual judgment. The 2022 amplitude_comp.png shows a ~1.6× uncalibrated gain error already happened once in this project's history.
  2. True-peak limiting below a fixed ceiling (default −1 dBTP).
  3. Level verification on the actual playback chain at session start, at the listener's normal listening volume, before any test material is played.
  4. Abort on NaN/Inf or discontinuity anywhere in the output buffer.

Rationale: the listener's residual acoustic hearing is irreplaceable and already degenerating. An over-level playback is an unacceptable, irreversible risk. This rule overrides experimental convenience, schedule, and every other consideration.


Tier 1 — Automated metrics (fast, every change)

These catch breakage and rank candidates before spending listening time. None of them is validated against CI percepts. Treat them as regression signals, not as truth.

1.1 Simulator correctness — the regression suite

The highest-value test available, and it costs no human listening time.

Validate against an external reference simulator whose parameters are documented. tests/test_reference_regression.py implements this; tests/fixtures/README.md explains how to supply matched source/reference pairs. The tests skip when fixtures are absent, so a clean checkout passes.

Compare in this order, the later measures mattering more than the earlier:

  • Per-channel envelope correlation — the primary measure, because envelopes are the information a real device transmits
  • Long-term average spectrum
  • Modulation spectrum

Exact waveform agreement is not expected: carrier phase and noise seeds differ by construction. Reference tools often output at a fixed sample rate regardless of input; resample before comparing.

Match a delta, not just an absolute. If reference outputs exist for the same source under two different configurations, verify that PRELUDE reproduces the direction and rough magnitude of the difference between them. That cancels implementation details irrelevant to the question and is a considerably stronger test.

1.2 Enhancement metrics

Computed on the electrodogram (channel × time stimulation matrix) rather than the waveform wherever possible — that is the actual information reaching the nerve.

Metric What it measures Caveat
Per-channel envelope correlation, CI(g(x)) vs x how much original structure survives the headline internal metric
Envelope modulation depth the CI's primary surviving cue maximizing blindly → pumping artifacts
Channel selection stability (n-of-m) how erratic the transmitted channel set is high instability suggests polyphony overload
Spectral contrast (peak-to-valley across channels) resistance to current-spread smearing
NCM (normalized covariance metric) envelope-based intelligibility among the better vocoded-speech predictors
ESTOI intelligibility envelope-based, reasonably CI-relevant
Onset-detection F1 vs. reference rhythm preservation rhythm is the CI's strength — protect it
Chroma / melodic contour agreement melody preservation pitch is the CI's weakness — the hard target
PESQ / POLQA designed for telephony codecs; poor CI fit. Use with suspicion or not at all.

Known failure mode: optimizing any of these directly produces audio that scores well and sounds worse. They rank candidates for human testing; they do not decide anything.

1.3 Hygiene checks

Output level within target; no clipping; duration preserved; sample rate and channel count as declared; config hash recorded in the artifact sidecar.


Tier 2 — Human listening sessions (the ground truth)

2.1 Two distinct experiment types — do not mix them

Type A — Simulator fitting. Does our simulation match the listener's implant percept? For bimodal listeners: present a candidate simulation to the acoustic ear only, and ask the listener to compare it against what the same source sounds like through the implanted ear. Where available, this is the only direct instrument for the measurement.

Type B — Enhancement preference. Is g(x) better than x through the implant? Present both through the implant path, blinded and level-matched, and ask for preference, intelligibility, or effort.

Type A calibrates the model. Type B tests the product. Conflating them produces uninterpretable data.

2.2 Session design rules

Derived from the elicitation caveats in 01-DOMAIN-PRIMER.md §4.1:

  1. Forced choice over free description. "Which of these two is closer to your CI?" beats "describe your CI." Free description is valuable for generating hypotheses, useless for testing them.
  2. Blind and randomise presentation order, by script rather than by hand. Investigators are typically invested in a positive result; where investigator and participant have a personal relationship, that investment is stronger and the bias is not hypothetical.
  3. Catch trials. Include identical A/A pairs to estimate response noise. Without a noise floor, a 60% preference is uninterpretable.
  4. Level-match everything (Tier 0).
  5. Short sessions. 15–20 minutes of active comparison. Auditory fatigue degrades discrimination, and the degraded data is indistinguishable from a null result.
  6. Fixed stimulus set across sessions for longitudinal comparability, plus a small rotating novel set to detect learning effects.
  7. Record the audiogram date alongside every session. In progressive loss the acoustic ear is a moving target, and a result from six months ago was measured on a different instrument.
  8. The participant may stop at any time, for any reason or none. Consent is ongoing and revocable, and this must be stated at the start of every session.

2.3 Stimulus ladder

Adopted from an earlier internal study design. The principle: isolate single perceptual dimensions before combining them, so that a failure can be localized rather than merely observed.

Tier Count Content Isolates
1 20–30 Pure tones across frequencies place/pitch coding, frequency resolution
2 20–30 Single instruments, diverse timbres spectral envelope, harmonic structure
3 10–20 Short passages, simple melodies melodic contour, F0 tracking over time
4 5–10 Complex polyphonic music n-of-m competition, source separation

Add a Tier 0: speech (audiobook excerpts) — the easiest CI case, the most likely daily use, and the tier where an early win is most achievable.

Ready-made resources worth seeking out:

  • Some published CI demonstrations ship speech/music/noise pre-rendered at 1, 4, and 16 channels — a channel-count ladder useful for establishing how many effective channels the listener has, independent of our own code.
  • the reference fixtures holds the parameter-documented act_/sim_ speech pair.

Rating dimensions: clarity, timbre accuracy, pitch perception, instrument separation, overall enjoyment. Sound decomposition — but collect them via forced-choice comparison rather than absolute 1–10 scales wherever the question permits (§2.2 rule 1).

2.4 What to capture per session

Logged in docs/lab-notebook/YYYY-MM-DD-session.md:

  • Date, time of day, the listener's subjective alertness/fatigue
  • Equipment: device, program/MAP slot, streaming path, playback volume setting
  • Any recent MAP change or audiogram change
  • Stimulus list with config hashes
  • Raw per-trial responses (not just summaries — the summary can be recomputed, the raw cannot be recovered)
  • Verbatim descriptive quotes — especially the vocabulary the listener uses for CI percepts. This is genuinely valuable primary data. There is no established vocabulary for describing electric hearing; theirs, accumulated over time, becomes the project's measurement language.
  • Anything surprising, including results that contradict the hypothesis

2.5 Priority: capture calibration data early

Type A experiments depend on residual acoustic hearing. Where that hearing is progressive, the window for collecting this data is finite and closing, whereas engineering work can be done at any time.

This inverts the intuitive ordering. It is worth running Type A sessions with whatever parameter-documented reference material is already available, before the software is finished, because those sessions cannot be run later. Software can wait; the measurement cannot.


Tier 3 — Longitudinal

  • Listening effort / fatigue over a long session (e.g. a full audiobook chapter) — arguably the most meaningful real-world outcome, and easier to detect than moment-to-moment preference.
  • Self-reported enjoyment over weeks of ordinary use.
  • Re-measure a fixed reference stimulus set at every audiogram change, to track the acoustic ear's decline and re-fit the bimodal crossover.

Reporting standard

Every experiment writes a lab-notebook entry stating: hypothesis, method, N, result, and an honest interpretation including negative results.

A negative result honestly recorded is a contribution. An overstated one is a harm to someone with a real personal stake in the outcome. Do not round a null result up.