判别式
突出
语音识别
计算机科学
发声
特征(语言学)
演讲制作
二元分类
任务(项目管理)
稳健性(进化)
人工智能
普通话
心理学
语音处理
听力学
倒谱
Mel倒谱
质量(理念)
隐马尔可夫模型
自然语言处理
医学
韵律
特征选择
作者
Hyunsun Ham,Jiwon Son,Keun You Kim,So Yeon Jeon,J Lee
标识
DOI:10.1093/geroni/igaf122.4206
摘要
Abstract This study investigates whether task-specific vocal acoustic features can reliably predict amyloid PET positivity in Alzheimer’s disease. While prior works often relied on a single task such as picture description, we systematically designed five distinct speech tasks, each designed at a different linguistic level—from acoustic-phonetic to discourse-level—to examine how the underlying pathology manifests in speech production at varying linguistic levels. Seventy-five participants completed five speech tasks, including 46 amyloid-negative and 29 amyloid-positive individuals, as determined by PET imaging. From each task, we extracted 162 hand-crafted acoustic features using an optimized clinical speech analysis pipeline. These features included spectral features(MFCCs), prosodic features (pitch, intensity, speech rate), and voice quality measures (jitter, shimmer, HNR). Recursive feature elimination and XGBoost were employed for feature selection, followed by task-specific classification using multiple binary classifiers. Among the tasks, the Sequential Motion Rate task yielded the highest accuracy(79%). Pitch-related features consistently ranked among the most important across all tasks, suggesting their robustness regardless of linguistic level. Prosodic features demonstrated increased discriminative power in higher-order tasks, whereas voice quality features were particularly salient in tasks emphasizing phonation and articulation. These findings highlight that vocal acoustic markers from specific speech tasks can effectively reveal amyloid burden, with distinct acoustic patterns depending on the linguistic complexity of the task. These task-dependent speech markers may reflect early neurodegenerative changes, supporting a link between amyloid pathology and speech production.
科研通智能强力驱动
Strongly Powered by AbleSci AI