Problem
Research questions and friction points this paper is trying to address.
Automatic Speech Recognition
distribution shift
robustness
continual learning
real-world conditions
MoDiCoL: A Modular Diagnostic Continual Learning Dataset for Robust Speech Recognition
🤖 AI Summary
Current automatic speech recognition systems exhibit insufficient robustness when confronted with real-world distribution shifts arising from the co-occurrence of diverse factors such as recording conditions, accents, speech impairments, and background noise—variables that existing datasets typically treat in isolation. This work proposes a dynamic perspective on robustness as an evolving capability and introduces MoDiCoL, the first modular diagnostic continual learning dataset designed to support multi-factor co-occurrence analysis. By constructing a realistic continual learning curriculum that systematically controls interactions among linguistic content, speaker characteristics, and acoustic environments, the study evaluates three continual learning strategies. The results elucidate how models acquire, transfer, and forget robustness under evolving conditions, offering new benchmarks and insights for developing reliable speech recognition systems.
📝 Abstract
Modern Automatic Speech Recognition (ASR) systems have made remarkable progress on standard benchmarks, yet performance gaps have emerged under real-world distribution shifts, caused by recording conditions, accents, speech impairments, and noise. Existing datasets and benchmarks typically isolate these factors, which overlooks their co-occurrence in real-world applications. In this paper, we argue that model robustness can be treated as a dynamic capability that continually develops, and we introduce MoDiCoL, a Modular Diagnostic Continual Learning dataset designed for controlled analysis of linguistic content, speaker characteristics, and acoustic environments. Furthermore, we propose a real-world-inspired continual learning curriculum to simulate incremental updates and study how robustness is acquired, transferred, and forgotten. We evaluate three continual learning strategies and provide detailed insights into robustness under evolving conditions.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Modular Dataset
Continual Learning
Robust Speech Recognition
Distribution Shift
Diagnostic Benchmark
🔎 Similar Papers
No similar papers found.
