Developing a Voice Recognition System for Data Entry of Gastrointestinal Endoscopic Findings to Meet the Demands in the IT Age

本研究发现，在日本，不同机构之间的实践差异和MST用户无法预测的树状结构阻碍了对内窥镜检查结果输入系统的接受。首先，我们构建了一个系统，其中考虑到了开发一个高自由度的日语内窥镜检查结果通用输入系统的重要性。具体来说，我们开发了一个内窥镜检查结果输入系统，通过启动MST作为普通内窥镜检查结果输入系统的背景术语，将MST术语与普通内窥镜检查结果绑定，而不是通过MST输入内窥镜检查结果。在语音输入系统方面，我们开发了按检查地点、症状等类别分层的分级词典，并动态选择每个词典的优化词汇表，以提高探索性处理效率和识别率。使用改进后的系统对实际朗读结果的识别率进行了评估。获得了不低于90%的识别率。在嘈杂的内窥镜检查场所进行的类似实验显示，识别率下降。我们的语音识别系统需要进行修改，以提高实际环境中的识别率。

The present study revealed that in Japan, the acceptance of the findings input system itself is hampered by variation in practice among different institutions and tree structures unpredictable by MST users.First, we constructed a system with "the importance of developing a versatile input system for endoscopic findings in Japanese at high freedom" in mind. More specifically, we developed an endoscopic findings input system that allows data entry without being conscious of MST, by booting MST as a background term for an ordinary endoscopic findings input system, with MST terms tied to ordinary endoscopic vocabulary, rather than by entering endoscopic findings via MST.Regarding the voice input system, we developed hierarchical word dictionaries stratified by examination site, symptom, and other categories, and dynamically chose a dictionary with the optimized vocabulary per dictionary, in an attempt to improve the exploratory processing efficiency and recognition rate. The recognition rate for actually-read-aloud findings was evaluated using the thus-improved system. A recognition rate of not less than 90% was obtained. Similar experiments at a noisy place for endoscopic examination revealed a decreased recognition rate. Our voice recognition system needs to be modified to improve the recognition rate in actual settings.