interspeech-writing-style

INTERSPEECH Writing Style

Four pages is not a short paper; it is a different literary form. The Interspeech

house style evolved to let a reader — who may be a phonetician, an ASR engineer,

or an area chair skimming forty PDFs — extract task, method, corpus, metric, and

delta in one pass. Everything below serves that one-pass property.

The opening contract

• Title: task and mechanism, in community vocabulary. Searchable in the ISCA

Archive beats clever.

• Abstract (~100–150 words): task → limitation of current practice → your move

→ corpus + metric + headline number → one-clause takeaway. Numbers belong in

Interspeech abstracts; "significant improvements" without a number reads as

evasion.

• Index Terms: the paper kit's index-terms line is your search surface — pick

the 3–5 terms your intended reviewers actually use.

• Introduction (~0.75 column): no history of deep learning. Name the task in

sentence one, the gap by sentence three, the contribution list by the end.

Density moves that buy space honestly

| Move | Before | After |

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

| Protocol one-liner | A paragraph on data prep | "We follow the ESPnet LibriSpeech recipe (commit X) except: …" |

| Diff-based method description | Re-derive the known architecture | "As in Conformer [n], with two changes: …" |

| Table consolidation | Separate tables per test set | One table, sets as column groups, best in bold |

| Citation compression | Three sentences of related work per item | Clause-level positioning: "unlike streaming approaches [4,5], we…" |

| Figure economy | Architecture diagram redrawing the standard | One figure that carries the finding (error breakdown, trade-off curve) |

The related-work section can shrink to a paragraph if positioning is woven into

the introduction — common and accepted at this venue (see

interspeech-related-work).

Speech-community diction

• Report metrics with their conditions attached: "WER 4.8% on test-other (no LM)",

"EER 0.92% on Vox1-O", "MOS 4.1 ± 0.08". A bare number is a red flag.

• Use canonical corpus/partition names; do not rename "dev-clean" to "validation".
• Past tense for what you did, present for what the system does, no "novel" as a

self-description — the reviewer decides novelty.

• Acronyms: expand once, then commit. An Interspeech paper drowning in undefined

acronyms fails the one-pass property for the scientist half of the audience.

• Claims scale with evidence verbs: "improves" needs a CI; "suggests" tolerates a

trend; "solves" needs more than four pages of anything.

The compression order (when you are at 5.5 pages)

1. Cut motivation prose — Interspeech readers arrive pre-motivated.
2. Convert method paragraphs to diffs against cited standards.
3. Merge tables; move per-condition detail into a results sentence.
4. Trim the intro's third rehearsal of the contribution.
5. Only then consider dropping an experiment — and if the paper still does not

fit, that is a Long Paper or journal signal, not a formatting problem.

Never buy space with template tricks; the PDF checker and the format police are

real (see interspeech-submission).

Micro-example: results sentence, before → after

Before (3 lines):  We evaluate our proposed method on the LibriSpeech
dataset. Our method achieves significant improvements over the baseline.
Results are shown in Table 2, which demonstrates the effectiveness.

After (2 lines):  On LibriSpeech, the proposed front end reduces WER from
5.6% to 4.9% on test-other (Table 2; paired bootstrap p<0.01), with the
largest gains on utterances under 3 s.

The after-version is shorter and carries corpus, metric, delta, significance,

and an analysis hook.

A working section budget for 4 pages

| Section | Budget | Job |

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

| Abstract + index terms | 0.15 col | task→move→corpus→metric→delta |

| 1. Introduction | 0.75 col | gap, hypothesis, contribution — no history lesson |

| 2. Relation to prior work | 0.3–0.5 col | clause-level positioning (may merge into 1) |

| 3. Method | 1.2 col | the diff against cited standards, one figure max |

| 4. Experimental setup | 0.6 col | corpus, protocol one-liners, disclosure block |

| 5. Results + analysis | 1.5 col | decisive table, ablation, one error-pattern finding |

| 6. Conclusions | 0.2 col | finding restated, scoped future sentence |

Budgets flex ±30% by paper type (science papers grow §4–5, systems papers §3),

but a draft violating the total by a full column will not compress overnight.

Bilingual/global audience note

Interspeech's readership is the most linguistically diverse of the major ML-adjacent

venues. Prefer plain constructions, keep sentences under ~25 words, and gloss

language-specific phenomena (tone sandhi, pitch accent) in one clause — reviewers

outside your language's community must still follow the claim.

Titles: a quick calibration

• Weak: "A Novel Deep Learning Framework for Improved Speech Processing" —

no task, no mechanism, unsearchable.

• Serviceable: "Prosody-Gated Chunking for Streaming ASR on Spontaneous Speech"

— task, mechanism, condition, all in Archive-searchable vocabulary.

• Community pattern: many memorable Interspeech titles name the system once

("Conformer: …", "SpecAugment: …") and then say plainly what it does; earn

the name after the colon with a descriptive clause, not instead of one.

Output format

[One-pass check] task/method/corpus/metric/delta extractable from p.1? y/n
[Opening] title, abstract-number, index-terms verdicts
[Density audit] paragraphs convertible to one-liners or diffs
[Diction] bare numbers, undefined acronyms, unverifiable verbs found
[Compression plan] ordered cuts to reach 4 pages without claim loss
[Escalation] fits in 4pp / needs Long Paper / needs journal

Format facts (4 content pages + references page; Long Paper track parameters)

follow the 2026 cycle as logged in resources/official-source-map.md

(2026-07-08); the paper kit of the current cycle is the sole format authority.