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
- Problem — the signal task, in its own vocabulary.
- Gap — why the standard method fails, with a specific named reason per prior line.
- Mechanism — the one thing you change, stated concretely.
- Measurement — the task-matched metric and where it improves.
- 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 /, mean over N runs" |
| "We use a deep neural network to..." |
"We replace the with , keeping " |
| "Excellent results across settings." |
Claim scoped to the SNR range and corpus actually tested |
| "State-of-the-art performance." |
"matches/beats under " |
| "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>
Source: brycewang-stanford/Awesome-Journal-Skills → ICASSP-Skills/skills/icassp-writing-style/SKILL.md
1---2name: icassp-writing-style3description: Use when revising an ICASSP paper into the IEEE two-column four-page house style — leading with the signal-processing contribution and mechanism, naming the task-matched metric on the first page, compressing method and results into 4+1 pages without a reviewed appendix, and writing math-forward claims that survive a subfield-expert reviewer.4---567# ICASSP Writing Style89Use this when revising the main paper. An ICASSP paper is a **four-page IEEE two-column object**10that must state a precise signal-processing contribution, show its mechanism, and prove it with a11task-matched number — all where a busy subfield reviewer will see it. There is no reviewed12appendix to rescue a buried argument.1314## Revision rules1516- Put the **signal-processing contribution and its mechanism** on the first page: the problem, why17 current methods break, what block of the signal chain you change, and the metric that proves it.18- Name the **task-matched metric** early (WER, SI-SDR, EER, PSNR, BER, RMSE — whatever the task19 uses), not a generic "accuracy" or "performance."20- Make the method **redrawable**: a reader should be able to reconstruct the block diagram or the21 estimator from the text alone.22- Use the four content pages for the core argument; push extended derivations and extra conditions23 to a cited technical report, since ICASSP has no reviewed supplement.24- Do not overclaim beyond the conditions tested; state the operating point (SNR, corpus, scene).2526## First-page arc27281. **Problem** — the signal task, in its own vocabulary.292. **Gap** — why the standard method fails, with a specific named reason per prior line.303. **Mechanism** — the one thing you change, stated concretely.314. **Measurement** — the task-matched metric and where it improves.325. **Why SPS cares** — the community relevance, in one sentence.3334## Compression discipline for four pages3536- One claim per paper; a four-page paper that hedges between two contributions lands neither.37- Every figure must earn its column-inches; a plot that does not change a decision is a paragraph38 you could have kept.39- Define notation once; two-column measure punishes redundant redefinition.40- Keep a one-line derivation sketch in the body for any result whose full proof lives off-paper.4142## Sentence-level rewrites4344| Draft pattern | ICASSP-safe rewrite |45|---|---|46| "Our method significantly outperforms prior work." | "reduces WER from A to B on <corpus>/<split>, mean over N runs" |47| "We use a deep neural network to..." | "We replace the <specific block> with <specific mechanism>, keeping <the rest>" |48| "Excellent results across settings." | Claim scoped to the SNR range and corpus actually tested |49| "State-of-the-art performance." | "matches/beats <named strong baseline> under <stated condition>" |50| "It is well known that..." | Cite the specific prior result you rely on |5152## Math-forward, but honest5354- State the signal and noise model before the estimator; a signal reviewer reads the model first.55- Keep problem-dependent constants and operating points visible rather than absorbed into vague56 "under mild conditions" language.57- Separate proved statements from empirically observed ones; mixing them in one paragraph reads as58 a credibility leak to an expert reviewer.5960## Vignette: compressing into four pages6162A draft with two methods, seven figures, and a long related-work section: keep the single stronger63contribution, its mechanism, and the two decision-critical figures (the metric-vs-condition curve64and the ablation); compress related work into explicit contrasts; move the second method and the65extra conditions to a cited report. The test of a good cut: a subfield reviewer can reconstruct the66whole argument without leaving the four pages.6768## Revision stub6970```text71[Writing diagnosis] clear / mechanism-hidden / overclaimed / overloaded72[First-page fix] <problem -> gap -> mechanism -> metric on page one>73[Metric fit] task-matched metric named early? yes/no74[Claim discipline] <claim -> figure/table/derivation, operating point stated>75[Compression cuts] <what moves off-paper to the cited report>76```7778## Output format7980```text81[Diagnosis] <one-line house-style verdict>82[Contribution sentence] <one signal-processing claim with its metric>83[Cuts] <move/merge/delete to fit 4+1>84[Overclaim risks] <phrases scoped to tested conditions>85```8687---8889**Source:** [`brycewang-stanford/Awesome-Journal-Skills`](https://github.com/brycewang-stanford/Awesome-Journal-Skills) → `ICASSP-Skills/skills/icassp-writing-style/SKILL.md`
thedixitjain