ICASSP Experiments

Use this before submission when the empirical story is not yet locked. ICASSP reviewers are subfield experts who know the right metric and the right baseline for your task, so the fastest route to rejection is the wrong ruler or a stale comparison. The four pages force a small number of decisive experiments, not a large number of weak ones.

Experiment audit

• Map each empirical claim to a specific table, figure, or condition sweep.
• Use the field-standard metric for the task; a novel or convenient metric invites the "that is not how this task is measured" review.
• Anchor to a current strong baseline and a standard corpus/benchmark, not to a weak or dated reference that flatters the result.
• Sweep the operating condition that matters (SNR, reverberation, bit rate, noise level); a single-condition number rarely convinces a signal reviewer.
• Report spread over runs (multiple seeds), and say in the caption whether bars are standard deviations, standard errors, or confidence intervals.
• Audit for train/test leakage, speaker/scene overlap across splits, and metric computed on the wrong crop, alignment, or normalization.

Match the metric to the task law

Task	Standard metric(s)	Standard evaluation anchor
Speech recognition	WER / CER	LibriSpeech, WSJ, or task corpus with fixed split
Enhancement / separation	SI-SDR, PESQ, STOI	Matched mixture set, reference-aligned scorer
Speaker / language ID	EER, minDCF	Standard trial lists (e.g., VoxCeleb-style)
Sound event / audio tagging	mAP, F1, error rate	Fixed labeled set, defined operating point
Image / video restoration	PSNR, SSIM	Standard test set, defined borders and depth
Communications	BER / BLER vs SNR	Defined channel model and decoder
Estimation / detection	RMSE, ROC/AUC	Monte-Carlo trials, bound (Cramér-Rao) if apt

Reporting the wrong metric family (e.g., classification accuracy for a separation paper) is a first-round reject pattern; match the ruler to the task before anything else.

What experiments are for at this venue

• ICASSP experiments exist to demonstrate a signal-processing mechanism works under realistic conditions, not to top a leaderboard by any margin. One clean condition sweep beats five extra datasets at a single point.
• The strongest design isolates the claimed mechanism with an ablation and shows it holds across the operating range, with the standard baseline drawn on the same axes.
• Where a theoretical bound exists (estimation, detection, coding), compare against it rather than only against another method.

Ablation and sweep stub

Fig. 2: metric vs condition (e.g., SI-SDR vs input SNR, 0-20 dB)
  - proposed (mean ± sd over 3 seeds)
  - strong baseline (same corpus, same scorer)
Table 1: ablation — remove one component at a time, same protocol
  - full method | -component A | -component B | baseline
Report: corpus + split, scorer config, seeds, run count, hardware/runtime

Vignette: a dereverberation paper

A submission claims improved dereverberation. The matching plan: evaluate on a standard reverberant set with PESQ and STOI using a fixed scorer, sweep reverberation time (RT60) rather than reporting one room, ablate the key module, draw a current strong baseline on the same axes, and report the mean and spread over seeds — every panel tied to the claim it supports.

Reporting floor

• Seeds and run counts for every stochastic figure; captions state what the error bars are.
• The actual compute and, for real-time claims, the measured latency or real-time factor — not a feasibility assertion.
• Honest disclosure of the condition you did not test, so a reviewer does not infer you hid it.

Output format

[Experiment readiness] strong / adequate / weak
[Metric fit] task-matched? <metric -> task>
[Baseline] current-strong / standard-corpus? yes/no
[Condition sweep] present over <axis>? yes/no
[Missing evidence] <ablation / spread / baseline / condition>
[Decision-critical next run] <one experiment>