icassp-writing-style

ICASSP Writing Style

Use this when revising the main paper. An ICASSP paper is a four-page IEEE two-column object

that must state a precise signal-processing contribution, show its mechanism, and prove it with a

task-matched number — all where a busy subfield reviewer will see it. There is no reviewed

appendix to rescue a buried argument.

Revision rules

• Put the signal-processing contribution and its mechanism on the first page: the problem, why

current methods break, what block of the signal chain you change, and the metric that proves it.

• Name the task-matched metric early (WER, SI-SDR, EER, PSNR, BER, RMSE — whatever the task

uses), not a generic "accuracy" or "performance."

• Make the method redrawable: a reader should be able to reconstruct the block diagram or the

estimator from the text alone.

• Use the four content pages for the core argument; push extended derivations and extra conditions

to a cited technical report, since ICASSP has no reviewed supplement.

• Do not overclaim beyond the conditions tested; state the operating point (SNR, corpus, scene).

First-page arc

1. Problem — the signal task, in its own vocabulary.
2. Gap — why the standard method fails, with a specific named reason per prior line.
3. Mechanism — the one thing you change, stated concretely.
4. Measurement — the task-matched metric and where it improves.
5. Why SPS cares — the community relevance, in one sentence.

Compression discipline for four pages

• One claim per paper; a four-page paper that hedges between two contributions lands neither.
• Every figure must earn its column-inches; a plot that does not change a decision is a paragraph

you could have kept.

• Define notation once; two-column measure punishes redundant redefinition.
• Keep a one-line derivation sketch in the body for any result whose full proof lives off-paper.

Sentence-level rewrites

| Draft pattern | ICASSP-safe rewrite |

|---|---|

| "Our method significantly outperforms prior work." | "reduces WER from A to B on <corpus>/<split>, mean over N runs" |

| "We use a deep neural network to..." | "We replace the <specific block> with <specific mechanism>, keeping <the rest>" |

| "Excellent results across settings." | Claim scoped to the SNR range and corpus actually tested |

| "State-of-the-art performance." | "matches/beats <named strong baseline> under <stated condition>" |

| "It is well known that..." | Cite the specific prior result you rely on |

Math-forward, but honest

• State the signal and noise model before the estimator; a signal reviewer reads the model first.
• Keep problem-dependent constants and operating points visible rather than absorbed into vague

"under mild conditions" language.

• Separate proved statements from empirically observed ones; mixing them in one paragraph reads as

a credibility leak to an expert reviewer.

Vignette: compressing into four pages

A draft with two methods, seven figures, and a long related-work section: keep the single stronger

contribution, its mechanism, and the two decision-critical figures (the metric-vs-condition curve

and the ablation); compress related work into explicit contrasts; move the second method and the

extra conditions to a cited report. The test of a good cut: a subfield reviewer can reconstruct the

whole argument without leaving the four pages.

Revision stub

[Writing diagnosis] clear / mechanism-hidden / overclaimed / overloaded
[First-page fix] <problem -> gap -> mechanism -> metric on page one>
[Metric fit] task-matched metric named early? yes/no
[Claim discipline] <claim -> figure/table/derivation, operating point stated>
[Compression cuts] <what moves off-paper to the cited report>

Output format

[Diagnosis] <one-line house-style verdict>
[Contribution sentence] <one signal-processing claim with its metric>
[Cuts] <move/merge/delete to fit 4+1>
[Overclaim risks] <phrases scoped to tested conditions>