Interspeech's supplement story is inverted from big-ML venues: there is no reviewed appendix. A regular paper is 4 pages of content plus one page for references and acknowledgments only — everything else is optional accompaniment that reviewers may open but never owe you. Design the paper to stand alone, then design the accompaniment to be irresistible. (Long Papers, new in 2026, follow their own reported 8+2 contract — verify before assuming appendix room.)
| Shelf | What lives there | Review status |
|---|---|---|
| The 4 pages | Every claim, the decisive table, the assumption set, protocol one-liners | Reviewed; the paper is this |
| References/acks page | Citations; acknowledgments (camera-ready only) | Policed — nothing else allowed |
| Accompaniment | Audio demo page, code repo, extended tables, media attachments | Optional reading; anonymity rules apply |
The classic failure: a decisive ablation "in the demo page" that a reviewer never
opens, followed by a review saying the claim is unsupported. If the paper's fate
depends on it, it goes in the 4 pages — compression is the skill (see
interspeech-writing-style), not relocation.
Interspeech is the one major venue where reviewers routinely listen. For synthesis, conversion, enhancement, and separation papers, the sample page is effectively part of the evidence:
interspeech-artifact-evaluation applies to every linked byte.Cycles differ on whether media files may be attached in CMT and what the size cap is — check the current author instructions rather than assuming. Attached media is frozen and safely anonymous; linked pages are richer but riskier (identity, mutability). When policy is unclear: attach the minimum decisive audio, link the rest, freeze both.
Where an AISTATS or NeurIPS author would write "App. C", an Interspeech author chooses:
RESULTS.md — full sweeps, per-language tables, negative results;
discoverable post-acceptance, invisible to review.Pick consciously per artifact; the default of "dump everything in the repo" leaves reviewable evidence outside the review.
Anonymous demo — Paper #1234
Section 1: Main comparison (Table 2 systems, 6 utterances × 4 systems)
utt-01 [GT] [Baseline] [Ablation] [Proposed] ← same grid every row
Section 2: Failure cases discussed in Sec 4.3 (3 utterances)
Section 3: Operating points (320/480/640 ms, 2 utterances each)
Selection: utterances drawn randomly from test; seed 17.
The grid mirrors the paper's tables, failure cases are volunteered rather than hidden, and the selection policy is printed on the page itself. A reviewer can audit the audio claims in under five minutes — which is the entire budget you realistically have.
audio_final_v2.wav names.[ ] Paper stands alone: no claim depends on unreviewed material
[ ] Sample page: same utterances across systems, failure cases included,
selection policy stated, mapped to table rows
[ ] Hosting anonymous, page frozen, media metadata scrubbed
[ ] Attachment rules of the current cycle checked (size, formats)
[ ] References page contains references (and, at camera-ready, acks) only
[ ] Extended-material route chosen: body / arXiv-later / repo / journal
[Shelf audit] what sits on each shelf; anything decisive off-page?
[Audio evidence] curation quality, selection policy, anonymity result
[Attachment plan] attach vs link, per current-cycle rules (cite source)
[Overflow route] compress / arXiv / repo / Long Paper / journal — per item
[Risk] <the unreviewed item most likely to be assumed reviewed>
Current-cycle attachment and preprint rules were checked 2026-07-08 against the
Interspeech 2026 author pages via renderings (resources/official-source-map.md);
the anonymity-period wording there governs anything you host.