5.2 KiB
Targeted ICASSP 2027 manuscript
This directory is an independent acceptance-target manuscript. It does not
modify or replace ../main/.
Positioning
The paper now starts from the egocentric conversational agent's when-to-speak problem, positions it against 27 cited works in turn-taking, first-person audio-visual conversation, active-speaker modeling, and streaming Omni interaction, then studies two cutoff-aligned interfaces:
- short-window encoder evidence (SWE): direct recent audio and optional person-crop visual encoder features alongside the frozen Omni readout;
- partial-tail refresh (PTR): query only the observed prefix of the current two-second block at each 200-ms cutoff, reusing completed payloads.
SWE names the short temporal support of encoder evidence, not a shallow network layer. The two-second chunk refers to this Qwen extraction setup, not a universal Omni-model setting.
The evidence remains a causal global--local audit, not a claim that all Omni models lack local perception or that PTR's independent gain is proven:
- global Omni context and recent audio are complementary across all horizons;
- most measured benefit is dense forecasting and wearer-offset anticipation;
- pooled and current-frame person vision do not establish speaker grounding;
- a three-seed causal oracle shows that timely visible-participant activity is informative, while identity continuity is not currently necessary.
The draft also reflects two seed-0 exploratory screens reported in
../main/main.tex: no-box full-frame F3 (Table 1) and a global-only
chunk-construction comparison (Table 4). Neither supplies
matched visual-motion evidence, endpoint gains, or the pending T/C factorial.
The result directories cited by the main draft for F3 and its G+A recheck
are absent from this checkout, so their underlying logs have not been
independently audited here.
The Results section also reports the archived 2026-08-24 sparse Thinker-LoRA diagnostic as a separate historical screen. Its strict causal-valid mean mAP declined under the recorded LoRA recipes, but the legacy feature/head contract is not matched to the current G+A matrix, and the remote raw metrics and checkpoints are unavailable here. It is not a Table 1 model ranking or proof that LoRA cannot help this task.
No semantic ENTRY/YIELD, addressee, online-detector, or real-time-throughput
claim is made. The deployable matrix is explicitly single-seed and exploratory.
EasyCom has archived screening scores. The Ego4D manifest and extraction /
training workflow exist, but this checkout has no archived Ego4D test metrics;
the manuscript specifies its transfer protocol and explicitly leaves any
cross-dataset performance claim pending.
The six-group L/T/C/B/H-Q/S table remains in main.tex as an
illustrative, unmeasured planning aid, excluded from the page-limited
PDF. Its directional statements are hypotheses and falsifiers, not EasyCom
scores. Replace or remove every row after the corresponding experiment,
including results that contradict the working story.
Build
cd manuscript/targeted
make
The editable sources are main.tex, tables_compact.tex, references.bib,
and figures/system_overview_onecolumn.tex. The earlier
figures/task_overview.tex is retained as unused source.
Generated LaTeX files are written to build/.
The preview targets four technical pages plus a reference-only fifth page.
The single-column method diagram (Fig. 1) has SWE on top and PTR below;
the earlier standalone task figure is not included.
The current PDF reserves seven numbered result tables. Table 1 spans two
columns and reports 0.2--2.0-s mAP for EasyCom and the pending Ego4D transfer
rows. Only three EasyCom models have archived complete horizon curves; other
intermediate cells remain unavailable. The old Table 1B oracle audit,
Table 2A T/C matrix, and Table 2B chunk screen are now standalone Tables
2--4. B (including an Ego4D transfer slot), H, and S occupy Tables 5--7.
Pending cells are explicitly unmeasured. Active table layouts are in
tables_compact.tex; the earlier full-precision tabulars
remain disabled in main.tex for source-level auditing.
Figure 2 is an H/Q case slot, not a selected or measured example. Its
cutoff-marked timelines must be inserted only after the slice protocol and
case-selection rule are frozen. The single-column system figure is generated
from figures/system_overview_onecolumn.tex; the earlier full-width source
is retained for comparison.
The build uses the repository's ICASSP 2027 LaTeX assets in
../ICASSP2027_Paper_Templates/, selected by latexmkrc. The supplied
spconf.sty and IEEEbib.bst match the 2026 copies.
Evidence sources
../../results/easycom_encoder_only_1s_4ep_2026-08-28/../../results/easycom_egospeak_three_level_eval_2026-09-16/../../results/easycom_omni_floor_screen_2026-09-16/../../results/easycom_oracle_pooled_wearer_audit_2026-09-10/../../results/easycom_sparse_supervision_vs_lora_2026-08-24/../../results/easycom_omni_lora_sparse_epoch2_2026-08-24/
The manuscript does not use superseded opportunity-conditioned ENTRY/INT
numbers. See REVIEWER_REPORT.md for the internal review loop and the remaining
scientific submission gates.