icassp-topic-selection

ICASSP Topic Selection

Use this before writing. ICASSP is the IEEE Signal Processing Society flagship, and its

defining feature is breadth: speech and audio, image and video, communications and radar,

sensor arrays, estimation and detection theory, and machine learning for signals all review under

one roof. Speech is one track among many, not the identity of the venue. A project fits when

its contribution changes a signal-processing primitive and can be proved in four pages.

Find the signal-processing primitive

Name what the contribution actually is:

• A new estimator, transform, filter, or detector; a new representation or embedding; a

new objective or inference procedure; a resource (corpus/benchmark); or a

model-based-deep-learning result that changes a block of the signal chain.

• If no signal-processing primitive exists — the paper is generic ML with no signal model — ICASSP

is likely the wrong venue (see the re-route table).

The ICASSP-vs-Interspeech routing decision (read this for any speech paper)

Speech papers can go to either venue, and picking wrong costs a cycle. They differ on mechanics as

much as on scope:

| Dimension | ICASSP (IEEE SPS) | Interspeech (ISCA) |

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

| Scope | All signal processing; speech is one track | Speech and spoken language only |

| Culture | Signal-processing methods, math-forward | Speech-science + engineering, spoken-language focus |

| Format | 4+1 (4 content pages + reference page) | 4+1 (4 content + reference page), different template |

| Anonymity | Single-blind — author list included | Double-anonymous with a pre-deadline anonymity period |

| Deadline | September (autumn) | Late February (winter) |

| Portal | CMS (cmsworkshops.com) | Microsoft CMT |

| Proceedings | IEEE Xplore | ISCA Archive, open access |

Decision rule: route to ICASSP when the contribution is a signal-processing method whose

speech application is one instance (a new front-end, estimator, separation objective, or array

technique), or when the September calendar and single-blind culture fit. Route to Interspeech

when the contribution is fundamentally about spoken language — phonetics, prosody, dialogue,

speech science, or a speech-specific model where the linguistic content is the point — or when the

February calendar and double-blind norms fit. When both fit, let deadline and reviewer community

decide.

Fit signal table (full ICASSP breadth)

| Signal in the project | ICASSP reading |

|---|---|

| New estimator/filter/detector with a signal model + evaluation | Core fit — the house genre |

| Speech/audio method framed as a signal-processing mechanism | Core fit (vs Interspeech if it is spoken-language-centric) |

| Image/video restoration, coding, or analysis | Core fit (compare ICIP for image-specific work) |

| Communications, radar, array, or sensor signal processing | Core fit |

| Generic deep net with strong benchmarks but no signal insight | Better at NeurIPS/ICML/ICLR/AAAI |

| A finished, journal-length result with extensive theory | An SPS journal (TSP/TASLP/SPL) |

Re-route table

• Generic ML, no signal model → NeurIPS, ICML, ICLR, AAAI.
• Spoken-language-centric speech → Interspeech (or SLT/ASRU for those subfields).
• Image-specific → ICIP; European/audio-workshop → EUSIPCO, WASPAA, DCASE.
• Journal-length, finished → IEEE TSP, TASLP, SPL, TIP, or a sibling SPS journal; consider the

OJSP-ICASSP track to present at ICASSP while publishing open-access in a journal.

Sharpening moves before committing

• State the primitive in one sentence; if you cannot, the ICASSP framing does not yet exist.
• Confirm the result fits four pages with a task-matched metric and a standard baseline; a claim

that needs journal-length exposition should go to an SPS journal or the OJSP track.

• For speech work, run the ICASSP-vs-Interspeech table explicitly rather than defaulting to

whichever deadline is closer.

• Subject-area emphasis drifts by cycle; scan the current EDICS list before final routing.

Output format

[Fit] strong ICASSP / possible ICASSP / better elsewhere
[Primitive] estimator / transform / representation / objective / resource / model-based-DL / none
[If speech] ICASSP vs Interspeech -> <decision + reason (scope/calendar/blinding)>
[Best venue] ICASSP / Interspeech / ICIP / EUSIPCO / WASPAA / SPS journal / ML venue
[Contribution sentence] <one signal-processing claim>
[Next action] <framing, experiment, or venue switch>