Levenslang leren voor automatische spraakherkenning

Automatic speech recognition (ASR) systems are widely used in applications such as voice assistants and transcription services, and to remain accurate and inclusive they must continually adapt to new speakers, accents, domains, and recording conditions. However, such adaptation often leads to catastrophic forgetting, where improving performance on new data degrades performance on previously learned conditions—fundamentally limiting the scalability of ASR systems to diverse and evolving user populations. Continual learning (CL) seeks to address this challenge by enabling models to extend their knowledge over time without sacrificing past abilities. This thesis advances CL for ASR with the goal of enabling speech-recognition models to continually enhance their recognition capability under realistic constraints, where storing and retraining on large amounts of past speech data is impractical or disallowed. We explore a range of approaches—including introducing task-specific parameters, weight-averaging strategies, memory-efficient and Hessian-based adaptation schemes, and online dual-model consolidation—that improve the balance between learning new information and retaining prior competence. Together, these contributions move ASR closer to practical, deployable continual learning, supporting models that can evolve with the users and environments they serve.