interspeech-related-work

INTERSPEECH Related Work

At Interspeech, related work is compressed into clauses (the 4-page format leaves no room for a survey), which raises the stakes: the few citations you make signal whether you know the field's actual genealogy. Speech has a citation culture with sharp edges — canonical systems, challenge baselines, and corpus papers each have one correct citation — and reviewers check.

The five lanes to cover

Missing lane 1 is the deadly one: a paper that cites only arXiv preprints while ignoring two prior Interspeech papers on the same task tells the area chair the authors never searched the Archive.

Search the field like a speech researcher

• ISCA Archive (www.isca-archive.org) is open access and searchable across all Interspeech years plus ISCA workshops — search it by task term before writing a single positioning sentence.
• Interspeech papers carry DOIs of the form 10.21437/Interspeech.<year>-<n>; cite the archival version, not the arXiv mirror, when both exist.
• Sweep the last ~3 editions of Interspeech and the last ICASSP for your task — the annual+annual rhythm means the state of the art moves twice a year.
• Check the relevant challenge's latest edition; at Interspeech, "the CHiME-8 baseline" is a shared reference point that locates your numbers instantly.

Venue attribution: the trap list

Speech's most-cited papers are scattered across venues, and misattributing them is a credibility wound. Verified anchors (see resources/exemplars/library.md for the checked list):

• Conformer, SpecAugment, ECAPA-TDNN, Tacotron, SUPERB — Interspeech papers (2020/2019/2020/2017/2021 respectively).
• LibriSpeech and x-vectors — ICASSP, not Interspeech.
• wav2vec 2.0 — NeurIPS; Whisper — ICML; Common Voice — LREC.
• HuBERT — a journal (TASLP) paper despite its conference-era fame.

When unsure, resolve the DOI before citing the venue. Never trust a BibTeX file scraped from arXiv for venue fields.

Positioning under compression

With one paragraph (or less) of space, use the clause pattern per contrast:

Streaming Conformer variants [3,4] reduce latency but hold WER on
test-other above 6%; adapter-based domain methods [5] recover accuracy
but require target-domain text. We keep the streaming constraint of [3]
while removing the text requirement of [5].

Each cited item gets a property and a limitation; your contribution is defined as the complement. Three such sentences outperform a half-page related-work section.

Self-citation under double-anonymity

• Third person, always: "extending the system of [7]" even when [7] is yours.
• If your prior system is unmistakably identifiable, cite it anyway — omitting the obvious predecessor is worse — but do not claim ownership.
• The anonymity period (from one month pre-deadline to decisions, per the 2025/2026 policy) also constrains when your own extended preprint may appear.

Corpus and tool citation etiquette

• Every dataset gets its paper cited (and license stated — see interspeech-artifact-evaluation); every toolkit that shaped results (ESPnet, Kaldi, SpeechBrain…) gets its citation.
• Metrics with a defining paper (PESQ, STOI, minDCF conventions) are cited on first use if any nonstandard choice is made.

BibTeX hygiene for a speech bibliography

@inproceedings{gulati20_interspeech,
  title     = {Conformer: Convolution-augmented Transformer
               for Speech Recognition},
  author    = {Gulati, Anmol and others},
  booktitle = {Proc. Interspeech 2020},
  pages     = {5036--5040},
  doi       = {10.21437/Interspeech.2020-3015}
}

• booktitle = {Proc. Interspeech <year>} is the community norm; do not write "INTERSPEECH: Annual Conference of..." variants that fragment your reference list's style.
• Keep the DOI field — it is how readers reach the Archive from the PDF.
• Purge arXiv eprint entries for papers that have archival versions; mixed duplicate entries (same paper cited twice via different keys) are a classic reference-page space leak under the 4+1 format.
• The references page is capacity-limited too: with roughly 20–30 entries fitting comfortably, every citation must be load-bearing. Cut courtesy citations before cutting positioning ones.

Output format

[Lane coverage] ISCA / IEEE-sibling / challenge / crossover / corpus — hit or gap
[Archive sweep] last-3-editions search done? missing prior art found
[Attribution check] venues verified via DOI for all load-bearing citations
[Positioning clauses] each contrast has property + limitation + complement
[Self-citation] third-person compliance; anonymity-period exposure
[Fix list] <ordered>

Archive URLs and the anonymity-period wording were checked 2026-07-08 (resources/official-source-map.md); challenge editions roll annually — re-sweep at writing time, not at submission time.