Letter to the Editor: Commentary on Feasibility Study of Parkinson’s Speech Disorder Evaluation With Pre-Trained Deep Learning Model for Speech-to-Text Analysis (Korean J Neurotrauma 2024;20:168-179) Authors: Kyung Hwan Kim Published: 2024 Dear Editor: I commend the authors for their efforts in exploring the feasibility of using a pre-trained Wav2Vec deep learning model for speech-to-text analysis in individuals with speech disorders related to Parkinson’s disease (PD).1,2) This study is noteworthy for leveraging deep learning to address speech impairments in PD, presenting promising potential for enhancing clinical diagnostic tools. However, I offer several constructive suggestions to enhance the clarity, methodological robustness, and practical applicability in future studies. First, while the authors analyzed baseline characteristics such as the Hoehn and Yahr scale and Unified PD Rating Scale (UPDRS) scores, the description of how these metrics influence speech-to-text accuracy is limited. Expanding on the relationship between disease severity and speech recognition accuracy could provide valuable insight into the model’s utility across different stages of PD progression. Second, the study briefly compares accuracy metrics between individuals with PD and healthy controls (HC) but does not delve deeper into how the model distinguishes nuanced speech features between the groups. Given that PD-related speech impairments often manifest subtly, additional analysis of phonetic and prosodic characteristics could be invaluable, particularly for early detection in clinical settings. The dataset used in this study included instances where the UPDRS speech-related score was not zero for HC participants and cases where it was rated as 1 for PD patients. Examining how the Wav2Vec model processes these edge cases could provide insights into its sensitivity and specificity. For example, if the model demonstrates greater accuracy than expert-rated scales in distinguishing HC from PD, its potential as a diagnostic tool would be significantly bolstered. Additionally, while the study acknowledges the challenges posed by language differences, this limitation could be explored further. Since the Wav2Vec model is primarily trained on English data, its application to languages with distinct phonetic structures, such as Korean, may require adaptations. Incorporating multilingual fine-tuning, or language-specific tokenizers could strengthen the model’s adaptability across diverse linguistic contexts. Recent research by Korean researchers on Wav2Vec2-based automatic speech recognition for the Korean language underscores the potential for such advancements. Follow-up studies leveraging these approaches would be a valuable contribution in the future. In conclusion, while this study presents promising findings, addressing these methodological considerations could enhance the model’s robustness and clinical utility. I hope these suggestions will prove useful for future research in this important area. Footnotes Funding: No funding was obtained for this study Conflict of Interest: The author has no financial conflicts of interest References