Saved in:
Bibliographic Details
Main Authors: Roll, Nathan, Bhalerao, Pranav, Bartelds, Martijn, Pawar, Arjun, Tatsumi, Yuka, Ogunremi, Tolulope, Shani, Chen, Graham, Calbert, Sumner, Meghan, Jurafsky, Dan
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2601.06972
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915722398007296
author Roll, Nathan
Bhalerao, Pranav
Bartelds, Martijn
Pawar, Arjun
Tatsumi, Yuka
Ogunremi, Tolulope
Shani, Chen
Graham, Calbert
Sumner, Meghan
Jurafsky, Dan
author_facet Roll, Nathan
Bhalerao, Pranav
Bartelds, Martijn
Pawar, Arjun
Tatsumi, Yuka
Ogunremi, Tolulope
Shani, Chen
Graham, Calbert
Sumner, Meghan
Jurafsky, Dan
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2601_06972
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Categorize Early, Integrate Late: Divergent Processing Strategies in Automatic Speech Recognition
Roll, Nathan
Bhalerao, Pranav
Bartelds, Martijn
Pawar, Arjun
Tatsumi, Yuka
Ogunremi, Tolulope
Shani, Chen
Graham, Calbert
Sumner, Meghan
Jurafsky, Dan
Computation and Language
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.
title Categorize Early, Integrate Late: Divergent Processing Strategies in Automatic Speech Recognition
topic Computation and Language
url https://arxiv.org/abs/2601.06972