ICASSP Artifact Evaluation
Use this for evidence packaging around ICASSP. There is no formal artifact-evaluation
committee or badge at ICASSP as of the 2026 cycle; artifacts are voluntary and their value is
in review credibility and post-publication impact. Because review is single-blind, a
repository can carry your name from day one — no anonymous mirror is required.
Artifact plan
- Decide what evidence supports the four-page claim: training and evaluation code, the exact
dataset splits or trial lists, the scoring script, model checkpoints, seeds, logs, and a
handful of qualitative samples (audio, spectrogram, image, or signal plot).
- Keep the decision-critical numbers reproducible from the released package; a reviewer who
cannot regenerate the headline metric discounts it.
- Ship a minimal reproduction map: environment, dependencies, hardware, commands, expected
outputs, runtime, seeds, and any known nondeterminism (GPU kernels, thread counts).
- For restricted corpora you cannot redistribute, provide enough provenance and preprocessing to
reproduce from the licensed source without violating the data-use terms.
- Because links can be public, put the repository URL in the paper and test it from a logged-out
browser before submission.
What ICASSP evidence reviewers open first
| Claim type |
First artifact inspected |
Common failure caught |
| Recognition/detection accuracy |
Scoring script + dataset split |
WER/error computed with an undocumented text-normalization or split |
| Enhancement / separation gain |
Reference-aligned SI-SDR/PESQ/STOI scorer |
Metric computed with mismatched reference or windowing |
| Speaker / biometric verification |
Trial list + EER/minDCF scorer |
Non-standard trials that inflate the score |
| Image/signal restoration |
PSNR/SSIM script + test set |
Metric on a different crop, bit depth, or borders |
| Real-time claim |
Latency/RTF measurement harness |
Feasibility asserted, never measured |
Signal-processing reviewers will rerun a scorer on a provided output far more readily than they
will retrain a model, so make the measurement path turnkey before polishing training code.
Worked vignette: packaging a separation result
A hypothetical submission reports source-separation gain on a two-speaker mixture set.
- Ship the mixture generation as one parameterized script (seed, SNR range, corpus source),
not as constants buried in a notebook, so a reviewer can regenerate the exact test mixtures.
- Include the reference-aligned SI-SDR scorer with its settings; separation numbers are
meaningless if the alignment and permutation handling differ.
- Emit result tables directly from logged outputs so the paper's numbers and the artifact's
numbers cannot drift apart.
- Provide five listenable example mixtures with their separated outputs; a reviewer often judges
perceptual quality from these before reading the metric table.
Turnkey scoring stub
# One command should reproduce the headline metric from released outputs.
python3 score.py \
--hyp outputs/dev.hyp \
--ref data/dev.ref \
--metric si-sdr \
--config configs/scoring.yaml # window, alignment, permutation policy pinned here
# Expected: dev SI-SDR = <value in Table 1> (mean over 3 seeds)
Calibration anchors
- Assume only the README and one entry script get opened; design the package so the top-level
README reproduces the main number in one command.
- Do not confuse a released toolkit with an artifact: ICASSP reviewers want the exact recipe that
produced this paper's numbers, not a general library dump.
- Formats and sizes for any uploaded supplementary media are cycle-specific; verify against the
current paper kit rather than past years.
Output format
[Artifact role] public release / demo samples / scoring package
[Contents] code / data-split / scorer / checkpoints / seeds / samples
[Reproduction level] turnkey / scripted / descriptive / weak
[Measurement risks] metric ruler / split / alignment / normalization
[Fixes before release] <ordered list>