@inproceedings{roll-etal-2026-categorize,
title = "Categorize Early, Integrate Late: Divergent Processing Strategies in Automatic Speech Recognition",
author = "Roll, Nathan and
Bhalerao, Pranav and
Bartelds, Martijn and
Pawar, Arjun and
Tatsumi, Yuka and
{\`O}g{\'u}nr{\`e}m{\'i}, Tol{\'u}lop{\'e} and
Shani, Chen and
Graham, Calbert and
Sumner, Meghan and
Jurafsky, Dan",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.2333/",
pages = "52997--53012",
ISBN = "979-8-89176-459-0",
abstract = "In speech language modeling, two architectures dominate the frontier: the Transformer and the Conformer. However, it remains unknown whether their comparable performance stems from convergent processing strategies or distinct architectural inductive biases. We introduce Architectural Fingerprinting, a probing framework that isolates the effect of architecture on representation, and apply it to a controlled suite of 24 pre-trained encoders (39M{--}3.3B parameters). Our analysis reveals divergent hierarchies: Conformers implement a ``Categorize Early'' strategy, resolving phoneme categories 29{\%} earlier in depth and speaker gender by 16{\%} depth. In contrast, Transformers ``Integrate Late,'' deferring phoneme, accent, and duration encoding to deep layers (49-57{\%}). These fingerprints suggest design heuristics: Conformers' front-loaded categorization may benefit low-latency streaming, while Transformers' deep integration may favor tasks requiring rich context and cross-utterance normalization."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="roll-etal-2026-categorize">
<titleInfo>
<title>Categorize Early, Integrate Late: Divergent Processing Strategies in Automatic Speech Recognition</title>
</titleInfo>
<name type="personal">
<namePart type="given">Nathan</namePart>
<namePart type="family">Roll</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pranav</namePart>
<namePart type="family">Bhalerao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Martijn</namePart>
<namePart type="family">Bartelds</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Arjun</namePart>
<namePart type="family">Pawar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuka</namePart>
<namePart type="family">Tatsumi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tolúlopé</namePart>
<namePart type="family">Ògúnrèmí</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chen</namePart>
<namePart type="family">Shani</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Calbert</namePart>
<namePart type="family">Graham</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Meghan</namePart>
<namePart type="family">Sumner</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dan</namePart>
<namePart type="family">Jurafsky</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>In speech language modeling, two architectures dominate the frontier: the Transformer and the Conformer. However, it remains unknown whether their comparable performance stems from convergent processing strategies or distinct architectural inductive biases. We introduce Architectural Fingerprinting, a probing framework that isolates the effect of architecture on representation, and apply it to a controlled suite of 24 pre-trained encoders (39M–3.3B parameters). Our analysis reveals divergent hierarchies: Conformers implement a “Categorize Early” strategy, resolving phoneme categories 29% earlier in depth and speaker gender by 16% depth. In contrast, Transformers “Integrate Late,” deferring phoneme, accent, and duration encoding to deep layers (49-57%). These fingerprints suggest design heuristics: Conformers’ front-loaded categorization may benefit low-latency streaming, while Transformers’ deep integration may favor tasks requiring rich context and cross-utterance normalization.</abstract>
<identifier type="citekey">roll-etal-2026-categorize</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2333/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>52997</start>
<end>53012</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Categorize Early, Integrate Late: Divergent Processing Strategies in Automatic Speech Recognition
%A Roll, Nathan
%A Bhalerao, Pranav
%A Bartelds, Martijn
%A Pawar, Arjun
%A Tatsumi, Yuka
%A Ògúnrèmí, Tolúlopé
%A Shani, Chen
%A Graham, Calbert
%A Sumner, Meghan
%A Jurafsky, Dan
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F roll-etal-2026-categorize
%X In speech language modeling, two architectures dominate the frontier: the Transformer and the Conformer. However, it remains unknown whether their comparable performance stems from convergent processing strategies or distinct architectural inductive biases. We introduce Architectural Fingerprinting, a probing framework that isolates the effect of architecture on representation, and apply it to a controlled suite of 24 pre-trained encoders (39M–3.3B parameters). Our analysis reveals divergent hierarchies: Conformers implement a “Categorize Early” strategy, resolving phoneme categories 29% earlier in depth and speaker gender by 16% depth. In contrast, Transformers “Integrate Late,” deferring phoneme, accent, and duration encoding to deep layers (49-57%). These fingerprints suggest design heuristics: Conformers’ front-loaded categorization may benefit low-latency streaming, while Transformers’ deep integration may favor tasks requiring rich context and cross-utterance normalization.
%U https://aclanthology.org/2026.emnlp-main.2333/
%P 52997-53012
Markdown (Informal)
[Categorize Early, Integrate Late: Divergent Processing Strategies in Automatic Speech Recognition](https://aclanthology.org/2026.emnlp-main.2333/) (Roll et al., EMNLP 2026)
ACL
- Nathan Roll, Pranav Bhalerao, Martijn Bartelds, Arjun Pawar, Yuka Tatsumi, Tolúlopé Ògúnrèmí, Chen Shani, Calbert Graham, Meghan Sumner, and Dan Jurafsky. 2026. Categorize Early, Integrate Late: Divergent Processing Strategies in Automatic Speech Recognition. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 52997–53012, Budapest, Hungary. Association for Computational Linguistics.