interspeech-artifact-evaluation

INTERSPEECH Artifact Evaluation

Interspeech has no badge-granting artifact committee; the artifact culture is

community-enforced instead. Reviewers listen to your samples, and post-publication

readers judge the paper by whether the recipe reproduces. This skill treats

artifacts in the two states an Interspeech cycle forces: **anonymous during review,

permanent after acceptance**. Check the current author instructions for what may be

attached versus linked — attachment rules vary by cycle.

The speech-specific artifact set

| Artifact | Review-time form | Post-acceptance form |

|---|---|---|

| Audio samples (TTS/VC/enhancement) | Anonymous static demo page or attached files | Permanent samples page linked in camera-ready |

| Code + training recipe | Anonymized repo (no usernames in history/CI) | Public repo, tagged at the paper's commit |

| Model checkpoints | Usually withheld (size, identity risk) | Hosted release with license and card |

| Data/corpus contribution | Described + license stated in paper | Archived with DOI, documented splits |

| Scoring/protocol scripts | In the anonymized repo | In the public repo — the piece most reused |

Anonymous demo pages without leaks

Audio demos are where Interspeech anonymity dies. Before the deadline, sweep:

• Hosting identity: GitHub Pages under a personal account, university subdomains,

and cloud buckets named after the lab all deanonymize. Use a fresh neutral host.

• Page furniture: analytics IDs, favicon, footer credits, CNAME records.
• The audio itself: a distinctive proprietary voice can identify the lab as surely

as a name — consider whether samples themselves are identifying.

• File metadata: WAV/MP3 tags, encoder strings, creation paths.
• Immutability: do not edit the page during review; a changing page suggests

post-deadline work and can carry timestamps.

Voice data is personal data

Speech artifacts carry legal and ethical weight that generic ML artifacts do not:

• Corpus licenses bind derivatives. VoxCeleb-trained speaker models, LDC-derived

lexicons, and YouTube-scraped audio each constrain what you may redistribute.

State, in the paper, the license of every corpus used.

• Speaker consent governs sample release: a cloned or converted voice needs the

source speaker's permission to be published, even "just for the demo page."

• Cloning/spoofing dual use: for synthesis and voice-conversion systems, say

what you release (recipe? checkpoint? nothing beyond samples?) and why. The

community norm — visible in anti-spoofing challenges like ASVspoof — is to weigh

release against misuse and to say so explicitly.

• Do not ship raw audio of human subjects unless its license explicitly allows

redistribution; ship feature extraction from the official corpus instead.

Minimum viable recipe

A speech result is reproducible when a stranger can regenerate the number, not

just run the model. Package for the number:

release/
├── README.md          # env, corpus versions + licenses, one command per table row
├── recipe/            # data prep → training → decoding, in run order
├── conf/              # exact configs/hyperparameters used in the paper
├── scoring/           # the WER/EER/MOS-analysis scripts and text-norm rules
├── RESULTS.md         # expected outputs with tolerances (seeds, CI width)
└── LICENSE            # code license + data-license pointers

The scoring/ directory matters most at this venue: WER moves with text

normalization and EER with trial lists, so publish the measurement, not only the

model (see interspeech-reproducibility).

Timing across the cycle

• Before submission: anonymous demo page frozen; repo anonymized if linked.
• During review: touch nothing linked from the paper.
• At camera-ready (19 June in the 2026 cycle): flip links to permanent public

homes; tag the repo; add the paper's ISCA Archive DOI to the README.

• Conference week: QR to samples on the poster; the demo page is your talk's

safety net when room audio fails.

Leak patterns seen in the wild

Concrete ways speech artifacts have deanonymized their authors — sweep for each:

• A demo page hosted at <university>.github.io/<lab-project>/ linked straight

from the anonymous PDF.

• Git history in a "fresh" anonymized repo carrying the original committer

emails (git log --format='%ae' | sort -u before sharing).

• WAV files whose RIFF metadata names the workstation user; MP3 ID3 tags naming

the recording engineer.

• A samples page reusing the lab's distinctive in-house TTS voice from previous

published demos — recognizable by ear.

• Cloud-bucket URLs containing the grant acronym.
• Analytics or font-loader scripts on the demo page tying it to the lab's other

properties.

Checkpoint release decision aid

| Situation | Community-normal release |

|---|---|

| ASR/SSL encoder on licensed-but-public corpora | Recipe + checkpoint |

| Speaker-verification model on VoxCeleb-family data | Recipe + checkpoint, license noted |

| TTS/VC on a consenting or synthetic voice | Recipe + checkpoint + samples |

| TTS/VC cloning a real speaker without release rights | Recipe + samples only; no checkpoint |

| Anything trained on scraped, unlicensed audio | Describe honestly; release nothing derived |

Output format

[Artifact inventory] samples / code / checkpoints / data / scoring — state of each
[Anonymity sweep] host, page, metadata, voice-identifiability findings
[License chain] corpus licenses → what may be redistributed
[Consent/dual-use] speaker permission status; release decision + rationale
[Release plan] what flips public at camera-ready, where it lives permanently
[Gaps] <ordered fixes>

Attachment size limits, link policies, and any new artifact requirements are

cycle-volatile — verify on the current author pages (2026 sources logged in

resources/official-source-map.md, checked 2026-07-08).