計畫成果與分析

This study explores personalized Mandarin ASR for the speech impaired. Sentence and short phrase tasks selected from the VoiceBank-2023 corpus are examined for amyotrophic lat-eral sclerosis and esophageal-speech speakers. Unadapted speaker-independent ASR model is trained with the NER-Pro corpus to serve as a pretrained model for the follow-up speaker-independent and dependent models fine-tuned with the CDSD and VoiceBank-2023 corpora. Several fine-tuning strategies are investigated. Experimental results show that the adapted speaker-dependent models attain average character error rates (CERs) ranging from 6.8% (normal) to 11.9% (mod-erate dysarthria) and 5.9% (normal) to 12.9% (moderate) for sentence and short phrase tasks, respectively. Last, we have implemented the fine-tuning strategies on a large corpus of a hereditary spastic paraplegia patient to evaluate frequently used sentence tasks under open-set and close-set conditions, resulting in CERs of 10.8% and 3.8%, respectively.

‍

The Truku language is one of the minority Indigenous languages of Taiwan. This paper presents the development of Taiwan’s first-ever Truku language text-to-speech (TTS) model, utilizing over 8 hours of high-quality audio recordings collected from professional male and female Truku speakers. The recordings were processed using VITS2 (Variational Inference for Text-to-Speech) technology to train a robust Truku-TTS model, which can accurately synthesize natural sounding spoken language from written Truku text. To facilitate the model training, we leveraged the computational power of the National Center for High-performance Computing (NCHC), we were able to significantly enhance model training efficiency and performance. Furthermore, we integrated Optical Character Recognition (OCR) technology into the Truku-TTS model workflow, allowing us to convert printed Truku language books into digital audiobooks. This innovative approach not only preserves Truku language but also broadens its accessibility, allowing a wider audience to engage with Truku content. The Truku-TTS model contributes substantially to ongoing efforts to preserve and revitalize Truku languages and provides a crucial tool for cultural and linguistic education. Our methodology and results demostrate a framework that can effectively preserve Truku language materials, enhance Truku language promotion services, and further extend these efforts ro cover all 16 Indigeous tribes comprising 42 dialects in Taiwan, achieving the goal of revitalizing and sustaining Indigenous languages.

‍

Frailty in older adults is caused by functional declines that result in unstable gait. This study analyzed gait in 24 frail and 22 non-frail older adults using acceleration and angular velocity signals from a wireless tri-axial inertial measurement unit (IMU). After noise was removed through Savitzky-Golay and Butterworth filters, gait features correlated with frailty were proposed and evaluated through normality tests and statistical analysis. To evaluate the frailty of older adults based on significant gait features derived from statistical analysis, the primary accuracy achieved is roughly around 84–89 % in k-nearest neighbor, support vector machine, and random forest models. To provide clinicians with a good tool for monitoring frailty and support preventive healthcare and aging-in-place strategies, we propose a gait-based detection system with an optimal feature extraction scheme that can exhaustively enumerate and evaluate potential parameters for optimal performance. This system significantly improved classification metrics (nearly all >95 %) with lower sensitivity and specificity and achieved 96 % accuracy with a portable, low-cost system that uses only one minute of walking data. These findings demonstrate that IMU-based gait analysis improves objectivity and accuracy in frailty classification. The optimal feature extraction scheme further refines performance, offering a scalable and time-efficient solution for community-based frailty detection. This approach highlights the potential of wearable sensors in improving geriatric health assessments.

‍

As the global average age increases and the number of patients with frailty and degenerative joint diseases like arthritis continues to rise, rehabilitation clinics often experience congestion due to the increasing demand for rehabilitation. Promoting proactive exercise among older adults and increasing their adherence can effectively reduce the prevalence of frailty. However, attracting users to regularly engage with and ensuring an easily usable operation flow and interface design will be crucial as an effective technological solution to improve the health issues of the elderly.This study proposes an upper limb game training software with a nostalgic retro theme, the game utilizes wearable devices with sEMG and g-sensor sensors for game interaction. Older adults will actually operate the game and complete the SUS usability questionnaire. The SUS questionnaire results for this game received an A-level usability rating. Based on the SUS scores, the upper limb game training software proposed in this study has development potential. The software is easy for the older adult population to engage with, regardless of age or past rehabilitation experience, and they have the willingness to continue using it.

‍

This study aims to explore the application of digital muscle strength testing systems in day care centers and analyze their adaptability and benefits for the elderly population. With the increasing global aging trend, sarcopenia has become a significant health issue, severely impacting the daily functioning and long-term care needs of older adults. Smart sensing technologies provide innovative solutions for long-term care facilities, enabling effective muscle strength and functional movement monitoring in elderly individuals. This study employs a mixed-methods approach, combining quantitative data analysis and qualitative interviews to assess the usability and acceptability of the intelligent testing system in long-term care institutions. The results indicate that the system performs well in muscle strength testing for older adults, with completion rates of 94.87% for the Dorsiflexion Ability and 92.31% for the 3M Gait Test. ANOVA analysis revealed no significant differences between the different stages of testing (3M Gait Test: F = 0.62, p = 0.54; 5 Times Sit-to-Stand Test: F = 0.62, p = 0.54). However, the results show a trend of “improvement at the mid-test and decline at the post-test,” highlighting potential challenges in the long-term implementation of the system in institutions. The qualitative analysis suggests that caregivers have a high level of acceptance of the system but also indicate areas for optimization in adapting the system to the “group nature” of the institution. This study provides valuable insights for optimizing and promoting digital muscle strength testing systems in long-term care facilities.

‍