A 115 MB adapter nearly matches full fine-tuning accuracy for dysarthric speech recognition.

Standard speech recognition often fails to understand people with dysarthria, a condition that makes speech hard to produce. Changing an entire speech model for every single patient takes up too much computer storage. Instead, researchers add small, plug-in adjustments called adapters to customize a system for just one speaker. The team tested seven adapter methods on two speech models for a post-stroke speaker with severe dysarthria. They discovered that basic LoRA adapters made huge improvements to transcription accuracy without needing complex upgrades. In fact, a popular version called DoRA offered no significant benefit over standard LoRA, while another variant, QLoRA, performed worse and saved no memory. Updating extra parts of the model came within 0.66 percentage points of changing the entire system, while using roughly 3.7 % of the storage. Furthermore, recording about 5 minutes of patient audio gave nearly half the accuracy gains of a 30-minute recording.

Why it matters

People with speech disorders like dysarthria often can't use standard voice assistants. This study finds efficient ways to personalize speech recognition for them, making technology more accessible without massive data or storage.

Questions answered