Categorize Early, Integrate Late: Divergent Processing Strategies in Automatic Speech Recognition
For researchers designing speech recognition architectures, this work provides empirical evidence of divergent processing strategies that can guide model selection for latency-sensitive vs. context-rich tasks.
The paper introduces Architectural Fingerprinting to compare Transformer and Conformer speech encoders, revealing that Conformers resolve phoneme categories 29% earlier and speaker gender 16% earlier in depth, while Transformers defer encoding to deeper layers (49-57%).
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.