interspeech-topic-selection

INTERSPEECH Topic Selection

Interspeech is the annual conference of ISCA — the **International Speech

Communication Association**, not to be confused with the computer-architecture

symposium that shares the acronym. It is the broadest speech venue in existence:

phoneticians, hearing scientists, ASR engineers, and voice-cloning teams review in

adjacent areas of the same program. Fit means convincing speech reviewers, and the

2026 theme in Sydney was "Speaking Together" — inclusion across languages, cultures,

and modalities. Reconfirm scope at https://interspeech2026.org before routing.

The modality test

Ask: if the audio were replaced by its transcript, would the contribution survive?

• If yes, the paper is NLP wearing a headset — route it to an ACL-family venue where

text-based reviewers will value it.

• If no — the claim depends on acoustics, prosody, speakers, articulation, hearing,

or the speech signal itself — Interspeech is a candidate.

• Borderline: spoken-language understanding and speech-LLM work fits when the spoken

channel (disfluency, ASR errors, prosody, audio tokens) is analyzed, not assumed away.

Where Interspeech sits among its neighbors

| If the core contribution is... | Prefer | Why not Interspeech |

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

| Speech science or technology, any subfield | Interspeech | — (this is home turf) |

| General signal processing, radar, imaging | ICASSP (IEEE) | Interspeech reviewers expect speech, not DSP breadth |

| ASR-only deep dive, workshop-scale | ASRU / SLT (IEEE, biennial) | Smaller, narrower; fine for follow-ups |

| Speaker/language recognition depth | Odyssey (ISCA workshop) | Specialist audience; Interspeech still fine |

| TTS system depth | SSW (ISCA Speech Synthesis Workshop) | Same trade-off as Odyssey |

| Text NLP, LLM reasoning, no audio analysis | ACL / EMNLP / NAACL | Fails the modality test |

| Mature work needing >4 pages of record | IEEE/ACM TASLP, Computer Speech & Language, Speech Communication | Journal timelines, no conference clock |

Subfield-to-area matching

Interspeech submissions are reviewed inside scientific areas chosen at submission

time. Misfiled papers draw reviewers with the wrong evaluation instincts:

• Recognition-side: ASR architectures, robustness, adaptation, decoding.
• Generation-side: TTS, voice conversion, singing synthesis.
• Talker-side: speaker verification, diarization, anti-spoofing, voice privacy.
• Human-side: phonetics, prosody, production/perception, pathological speech,

child speech, second-language learning.

• Interaction-side: spoken dialogue, SLU, speech translation, audio LLMs.
• Resource-side: corpora, benchmarks, evaluation methodology, challenges.

A phonetics claim reviewed by ASR engineers (or vice versa) is the most common

self-inflicted rejection at this venue; pick the area for the claim, not the tools.

Track choice inside Interspeech (new in 2026)

Interspeech 2026 introduced a Long Paper track (reported: up to 8 pages plus 2

reference pages, targeted acceptance below 30%) alongside the classic 4-page regular

paper. Choose long only when the contribution genuinely cannot be stated in four

pages — a benchmark with extensive analysis, a theory-plus-system arc. A padded

4-page idea in the long track competes against a stricter bar. Verify the current

track rules; this split is one cycle old.

Signals the project is Interspeech-shaped

• The evaluation uses speech-native metrics (WER, MOS, EER, PESQ) on named corpora.
• Error analysis engages speakers, accents, noise conditions, or phonetic categories.
• The related work traces through the ISCA Archive, not only arXiv and ACL.
• A speech scientist and a speech engineer would both find something to argue with.

Signals to reroute

• Results reported only on text benchmarks with audio as an afterthought.
• The novelty is a general ML trick demonstrated on speech among other domains —

NeurIPS/ICML reviewers reward generality; Interspeech reviewers ask what it

teaches about speech.

• The paper is a product description without a falsifiable claim.

Three triage vignettes

• "We fine-tuned an LLM to summarize meeting transcripts." Transcript-only

input, text output, no acoustic analysis — fails the modality test. Route to an

ACL-family venue; return here only if ASR-error propagation or prosodic cues

enter the analysis.

• "Our vocoder halves inference cost at equal MOS." Speech-native claim,

speech-native metric, engineering contribution — strong Interspeech fit in a

generation-side area; SSW is the specialist fallback if the delta is

vocoder-internals deep.

• "A cross-linguistic study of vowel reduction in child speech." No system at

all, and still a strong fit — human-side areas review speech science on its

own terms. This breadth is exactly what ICASSP does not offer.

Questions to settle before committing

• Which two or three scientific areas could plausibly host the claim, and which

reviewer pool do you actually want? (Decide now; it is frozen at submission.)

• Is there a running challenge whose test set and baseline make your evaluation

legible for free?

• Does the evidence you can gather by the deadline fill 4 pages convincingly, or

does honesty require the Long Paper track or a journal?

• If the answer is "reroute," is the finding still worth an ISCA workshop

(Odyssey, SSW, task workshops) where the specialist audience is denser?

Output format

[Modality test] survives transcript-only? yes / no / partial — <one line>
[Interspeech fit] strong / plausible / reroute
[Area match] <scientific area the claim belongs to>
[Track] 4-page regular / Long Paper (verify current rules) / workshop or challenge
[Reroute target] <ICASSP / ASRU / SLT / Odyssey / SSW / ACL-family / journal + reason>
[Cycle reality] next deadline and source URL, or 待核实

Cycle anchor (checked 2026-07-08)

Interspeech 2026: Sydney, Australia, 27 September – 1 October 2026, hosted with

ASSTA; submissions closed 25 February 2026. Interspeech 2027 is announced for São

Paulo, Brazil (29 August – 2 September 2027) — the first South American edition —

so a project failing today's fit test has a concrete next target. Treat every date

here as a snapshot; reopen the official site before planning.