Deep-Learning-Based Segmentation of Predefined Chunks in Connected Speech: A Retrospective Analysis

Abstract

Background and Objectives In institutional settings, manually segmenting connected speech is a time-consuming and labor-intensive process. This study aims to develop a deep-learning model for automating this process, evaluating its accuracy, and determining the minimum dataset size for effective performance.Materials and Method Voice data from 524 individuals with pathological conditions and 502 individuals with normal conditions, totaling 1026 samples, were used. Each voice sample had 17 chunks, including a “summer” sentence (15 chunks) and vowels /α/ and /i/. The deep-learning model employed in this study is based on the multi-layer perceptron-mixer architecture. This study evaluated performance using the Intersection over Union (IoU) metric, commonly employed in artificial intelligence-based image detection for chunk segmentation.Results The accuracy of chunk identification at the frame level was 96.47%. Using IoU metrics, chunk segmentation accuracy was 98.15% at IoU ≥0.6, 96.03% at IoU ≥0.7, and 89.78% at IoU ≥0.8. Optimal dataset size exploration indicated that more than 700 connected speech datasets were needed for successful training, maintaining F1-scores up to 95% at IoU ≥0.7.Conclusion The artificial intelligence model is suitable for the development of an automated system that efficiently divides segments in the institutional collection of voice data. This suggests its potential utility in advancing voice research using connected speech.