icassp-supplementary

ICASSP Supplementary

Use this when assembling everything around an ICASSP paper that does not fit the four content

pages. The defining constraint: ICASSP has **no reviewed supplement or appendix that decides

acceptance**. The paper must be self-contained in 4+1 pages; anything else is optional material a

reviewer may consult but is not obliged to.

The three shelves

| Shelf | What lives here | Reviewer obligation |

|---|---|---|

| Pages 1-4 | The claim, mechanism, and decision-critical results and figures | Must read |

| Page 5 | References, funding acknowledgements, ethical-standards statement only | Must read (references) |

| Optional release | Extended derivations, extra conditions, demo media, code | May consult, not guaranteed |

If a result decides acceptance, it belongs on pages 1-4. If it merely strengthens or illustrates,

it can live in the optional release — but never assume a reviewer opened it.

What "supplementary" means at ICASSP

• Unlike double-blind ML venues with a reviewed appendix, ICASSP's optional material is usually a

public repository or a demo page, and because review is single-blind it can carry your

identity from the start.

• Some cycles allow uploading a multimedia file (audio, image, or video) with the submission

through CMS; confirm the current paper kit for accepted formats and size limits before relying

on it.

• Long proofs and derivations that will not fit four pages typically go to an **extended

arXiv/technical-report version** cited from the paper, not to a reviewed appendix.

Curating multimedia reviewers actually open

For enhancement, separation, synthesis, restoration, or any perceptual result, a short set of

listenable or viewable examples is often the most persuasive artifact:

• Provide a small, curated set (roughly five), not a dump; pick cases that show both a strength

and an honest hard case.

• Label each example with its condition (SNR, corpus, or scene) so a reviewer can interpret it.
• For audio, include the input, the output, and a strong-baseline output side by side.
• Scrub embedded metadata only for privacy/consent reasons, not anonymity — single-blind review

does not require hiding your identity.

Keeping the four pages self-contained

• State each main theorem, algorithm, or system so a reviewer can judge it without the external

material; a paper whose argument only closes in the arXiv version reads as unreviewable in four

pages.

• Keep one derivation-sketch line in the body for any result whose full proof you exile to the

technical report.

• Cross-reference the release from the paper at least once; unlinked material is invisible.

Vignette: splitting an image-restoration paper

A submission proposes a denoiser with a stability argument and a broad evaluation. The body keeps

the method, the key stability statement with a one-line justification, and the two

decision-critical PSNR/SSIM tables; the extended derivation and the full noise-level sweep go to a

cited technical report; a demo page shows five before/after crops with a baseline. Nothing that

decides acceptance lives only off-paper, because ICASSP reviewers are not required to leave the

PDF.

Consent and licensing checklist

[On-paper] claim, mechanism, decision-critical results all in 4+1
[Off-paper] derivations / extra conditions / media / code (cited from paper)
[Media] curated set, labeled conditions, baseline shown
[Licensing] corpus/data license permits redistribution of any shipped samples
[Consent] speaker/subject consent cleared for any released recordings or images

Output format

[Supplement plan] ready / needs fixes / not ready
[Page budget] 4 content + 1 reference/acks/ethics respected?
[Self-containment] does the paper close without the external material?
[Media] curated / labeled / baseline-anchored
[Release link] present in paper and reachable