Variational Low-Rank Adaptation for Personalized Impaired Speech Recognition

Researchers developed a novel Bayesian Low-rank Adaptation method for personalizing automatic speech recognition systems to better understand impaired speech. The approach addresses challenges in ASR systems like Whisper that struggle with non-normative speech patterns from conditions like cerebral palsy, using data-efficient fine-tuning on English and German datasets.

• →Current state-of-the-art ASR models like Whisper perform poorly on speech from individuals with congenital disorders or acquired brain injuries.
• →Collecting and annotating impaired speech data is particularly challenging due to the effort required from affected individuals and need for familiar caregivers.
• →The new Bayesian Low-rank Adaptation method enables data-efficient personalization of ASR systems for impaired speech.
• →Validation was conducted on English UA-Speech dataset and a newly collected German BF-Sprache dataset from a child with speech impairment.
• →The approach significantly improves ASR accuracy while maintaining efficiency in low-resource settings.