Design of a Multiscale Temporal Aware Model for Evaluating Spoken Fluency in Oral Expression

Automatic fluency assessment is a critical task for real-time speech applications, such as language learning platforms and broadcast monitoring systems. This paper proposes a novel approach for evaluating spoken fluency by integrating Multiscale Temporal Convolutional Networks (TCN) with self-attention mechanisms. The model is designed to capture both local prosodic features and long-range temporal dependencies in speech, which are essential for accurate fluency scoring. The architecture incorporates multiscale TCN for extracting speech rhythm patterns at multiple temporal resolutions, and self-attention to model global dependencies across the entire sequence, enabling the model to dynamically focus on critical speech features such as pauses, speech rate, and pitch fluctuations. We evaluate the model on a newly constructed Broadcast Fluency Evaluation Corpus (BFE-Corpus), and compare it with several baseline models, including LSTM, CNN-LSTM, and Transformer. The experimental results demonstrate that our approach significantly outperforms baseline models in terms of Pearson correlation (0.842) and Root Mean Squared Error (0.738), while maintaining low inference time (6.1 ms) suitable for real-time deployment. We also demonstrate the model's generalization ability across different speakers and its performance on edge devices like Jetson Nano and mobile platforms, highlighting its potential for practical applications. The proposed model not only achieves a balance between accuracy and computational efficiency but also lays the groundwork for future developments in multimodal fluency assessment and personalized learning systems.