--- name: interspeech-workflow description: Use when managing an INTERSPEECH project across the annual ISCA cycle — backward-planning from the late-February deadline through the update window, spring review, June notification, camera-ready and registration coupling, and the September–October conference, including what to do in the current post-decision phase and 2027 planning. --- # INTERSPEECH Workflow Interspeech runs one cycle per year with an unforgiving rhythm: a winter deadline, a silent spring, a June verdict, and a Northern-autumn conference. This skill is the project-management layer; every date below is the verified 2026 (Sydney) calendar and must be swapped for the live one each cycle. ## The 2026 reference clock | Date (2026) | Event | Owner's job | |---|---|---| | 17 Jan | CMT portal opens | metadata drafted, area chosen | | 25 Feb | Paper deadline | PDF + frozen metadata in | | 4 Mar | Paper-update close | sanctioned typo/figure pass done | | ~24 Apr – 1 May | Rebuttal window (reported, 待核实) | response team on call | | 5 Jun | Notification | triage meeting same week | | 19 Jun | Camera-ready close | de-anonymized PDF + metadata | | 15 Jul | Early-bird registration closes | author registration secured | | 16 Aug | Standard registration closes | stragglers covered | | 27 Sep – 1 Oct | Conference, Sydney | talk/poster + audio fallback ready | ## Where the cycle stands today (2026-07-08) Mid-cycle, post-decision. Accepted teams: camera-ready is behind you; **early-bird registration closes in one week** — that is the live deadline — then presentation prep, artifact publication (`interspeech-camera-ready`, `interspeech-artifact-evaluation`), and travel/visa for Sydney. Rejected or unsubmitted teams: the 2027 cycle (São Paulo, 29 Aug – 2 Sep 2027) is the target; expect a deadline in the same late-winter region but **treat the 2027 dates as unpublished until the São Paulo site posts them**. ## Backward plan from any Interspeech deadline Offsets that have survived contact with real teams; anchor them to live dates: - **T−10 weeks**: claims frozen; corpus/split/licensing decisions final. - **T−6 weeks**: main results table filled with *some* number in every cell — the quality improves later, the design cannot. - **T−4 weeks**: draft exists in the official paper kit (not a one-column doc — the 4-page reality only shows in the template). - **T−2 weeks**: internal review by one speech scientist *and* one engineer, mirroring the reviewer mix; significance tests run. - **T−1 week**: compression pass (`interspeech-writing-style`), anonymity sweep, demo page frozen. - **T**: submit half a day early; CMT queues at AoE cutoffs. - **T+1 week**: proofread the submitted PDF; use the update window for repairs only. ## Failure patterns by phase - **Winter**: discovering at T−1 that the paper is 5.5 pages — compression is a week of work, not a night (see the compression order in `interspeech-writing-style`). - **Spring silence**: team disbands; nobody owns the rebuttal when it lands. Assign the response owner at submission time. - **June**: acceptance euphoria eats the two-week camera-ready window; the registration coupling (no registered author → no proceedings entry) surfaces in August as a crisis. - **September**: audio demos that only ever ran on the lab machine meet a conference-hall AV chain. Rehearse playback; carry spectrogram fallbacks. ## Parallel-track decisions - **Interspeech vs ICASSP double rhythm**: the sibling deadline (typically autumn for ICASSP) means a rejected February idea has a fall retry — decide in June whether reviews point to "fix and ICASSP" or "grow and Interspeech 2027". - **Challenge alignment**: many Interspeech papers ride a challenge (CHiME, VoxSRC, ASVspoof…); challenge timelines are their own clocks, published by each organizer, and system-description papers have separate rules — verify per challenge. - **Journal extension**: an accepted 4-pager with leftover material has a standard afterlife at TASLP / Computer Speech & Language / Speech Communication; start the extension outline while the reviews are fresh. ## Standing owner map Assign at kickoff, not at crisis: - **Paper owner** — format, compression, the submitted PDF, the update-window proofread. - **Evidence owner** — experiment matrix, seeds, significance, scoring scripts. - **Compliance owner** — anonymity sweep (PDF + demo page + repo), license chain, policy checkboxes. - **Rebuttal watcher** — has CMT credentials, watches the reported window, owns the response draft calendar. - **Logistics owner** — registration before the tier cutoffs, visa timelines, AV rehearsal for the talk or poster QR codes. One person can hold two hats; no hat may be unassigned. Every 2026 failure pattern above maps to exactly one vacant hat. ## Minimal milestone ledger ```text [frozen] claims / corpus / splits ........ T-10w owner: evidence [in-kit] draft in official template ...... T-4w owner: paper [reviewed] internal science+eng pass ....... T-2w owner: paper [swept] anonymity + demo page frozen .... T-1w owner: compliance [in] CMT submission confirmed ........ T owner: paper [patched] update-window proofread ......... T+1w owner: paper [watched] rebuttal window covered ......... spring owner: watcher [final] camera-ready + registration ..... June owner: logistics ``` ## Output format ```text [Phase] pre-deadline / update-window / silent-spring / rebuttal / post-decision / conference-run-up / post-conference [Live deadline] [Owner map] paper / rebuttal / camera-ready+registration / artifacts / talk [Backward plan] next three offsets with dates [Parallel tracks] ICASSP retry? challenge clock? journal extension? [Risk now] ``` Verified anchors and the 待核实 list live in `resources/official-source-map.md` (checked 2026-07-08). ISCA rotates organizing committees per edition — the Sydney team's choices (tracks, rebuttal shape, attachment rules) do not bind São Paulo.