brycewang-stanford INTERSPEECH Author Response
Interspeech added an author rebuttal to its historically response-free pipeline in recent cycles; for 2026 the response window was reported as 24 April – 1 May, ahead of the 5 June notification. Confirm the live mechanics (length cap, CMT form fields, whether replies go to reviewers or only the area chair) before writing a word — rebuttal formats at this venue are young and still moving.
Know your audience mix
An Interspeech review set typically mixes reviewer species, because ISCA's areas span science and engineering:
| Reviewer instinct | What they attacked | What convinces them |
|---|---|---|
| Speech engineer | Baselines stale, WER delta within noise | Point to the significance test or per-condition breakdown already in the paper |
| Speech scientist | No phonetic/error analysis, claims exceed perception data | Cite the analysis subsection; concede scope honestly |
| ML generalist | "Incremental over the arXiv version of X" | State the technical delta in one sentence; note what the 4-page record shows |
| Resource skeptic | Corpus license, split leakage, evaluation protocol | Quote the exact split/protocol line; name the license |
Answer each reviewer in their own dialect. A MOS-methodology objection answered with more WER numbers reads as a dodge.
The 4-page constraint shapes the rebuttal
Unlike venues with appendices, everything reviewable was on those four pages — so "see appendix" is never available. Your levers are:
- Locate: most objections at Interspeech are density casualties; the answer is in the paper but compressed. Quote section and line.
- Reframe: restate what the 4 pages could not spell out — one derivation, one clarified protocol — briefly, as clarification, not as new content.
- Concede and scope: if the evidence truly is not there, say so and state what a camera-ready (or the next cycle) would add. Speech reviewers punish bluster more than gaps.
Whether new numbers are allowed in the response varies by cycle; if unverified, assume tables of new results are unwelcome and offer them as camera-ready additions.
Anonymity does not pause
The response is part of the double-anonymous record and the anonymity period runs until decisions. No identity hints, no links to demo pages created after submission under identifiable accounts, no "as our group showed at Interspeech 2024."
Reply skeleton
One block per objection, ordered by decision weight:
R1-Q2 (statistical validity of the 0.4 WER gain):
- The concern: gains may be within run-to-run variance.
- In the record: Table 2 reports mean±sd over 3 seeds; Sec 4.2 gives the
bootstrap 95% CI on test-other, excluding zero.
- Clarification: the CI uses utterance-level resampling (1000 draws).
- Offered for camera-ready: add the matched-pairs test on test-clean.
Four to six lines each. The meta-reviewer will compress you anyway — pre-compress.
Triage order under a one-week window
- Day 1: read all reviews twice; tag each point evidence / clarity / scope / taste. Ignore taste unless the meta-reviewer could echo it.
- Days 2–3: draft the evidence answers first — they decide acceptance at a venue where "not convincing" is the modal reject phrase.
- Day 4: cross-reviewer consistency pass; contradictory replies to two reviewers are visible to the area chair on CMT.
- Day 5: cut a third; verify the character/word cap; submit early. CMT under deadline load is not where you want to be at 23:58 AoE.
What to skip
- Do not re-argue the venue ("this is more of an ICASSP concern") — the routing decision was yours at area selection.
- Do not promise a Long Paper resubmission as a fix for missing evidence.
- Do not thank at length; the cap is better spent on the CI you computed.
- Do not respond to score, only to content — scores at Interspeech shift through the meta-review, not through indignation.
Pre-send checklist
- Every objection from every reviewer is acknowledged, even if only with one scoped sentence — silence on a point reads as concession.
- Each factual reply cites a section, table, or figure of the submitted PDF.
- No new result tables unless the live instructions explicitly permit them.
- Replies to different reviewers tell one consistent story.
- Zero identity signals: no lab names, grant codes, fresh links, or "our Interspeech 2024 system" phrasing.
- Word/character cap verified against the CMT form, not against habit.
- Someone outside the author list read it cold and could restate each objection from your answer alone.
Emotional discipline, briefly
Interspeech reviews arrive compressed and sometimes blunt — three reviewers covering a 4-page paper leave little space for pleasantries. The productive reading: every blunt sentence is a compression artifact of a real concern. Extract the concern, drop the tone, and never mirror bluntness back; the meta-reviewer reads your temperament as evidence about the work's solidity.
If there is no rebuttal this cycle
Rebuttal mechanics at Interspeech are recent and not guaranteed to recur. If the
current cycle runs without one, this skill still applies at two later moments:
the camera-ready revision memo (accepted papers), where reviewer concerns get
addressed in print, and the next-cycle revision plan (rejected papers), where
the same objection map drives the rewrite — see interspeech-review-process
for the triage lanes.
Output format
[Review set] R1/R2/R3(+) one-line objection map, tagged by type
[Decisive objection] <the one the meta-reviewer will weigh>
[Anchors] <section/table/figure of the 4-page record answering each point>
[Concessions] <honest gaps + camera-ready offers>
[Rebuttal draft] <numbered blocks per skeleton, within the live cap>
[Verify first] cap / new-results policy / who reads it — from the live CMT form
The 2026 window dates above came from secondary renderings and are flagged 待核实 in
resources/official-source-map.md; the official author pages control.