Precise Guidance for English Prosody Perception Pronunciation Based on Deep Learning

Abstract

In English pronunciation teaching, prosody has long received less attention than segmental phoneme training, although stress, rhythm, and intonation strongly affect speech intelligibility and naturalness. Existing technological aids mainly focus on phoneme-level correction and still lack fine-grained diagnosis and adaptive feedback for prosodic waveform patterns. From an engineering perspective, prosody perception can be regarded as a speech-signal analysis problem involving pitch trajectory, energy distribution, temporal modulation, and waveform feature recognition, which is methodologically related to time-frequency analysis in wave-based signal processing. This study constructs an intelligent perception and precise guidance system for English prosody based on deep learning. Speech samples were collected from English learners with different dialect backgrounds, and a specialized corpus containing 4592 valid speech samples was constructed. Acoustic features including fundamental frequency, duration, intensity, energy rising points, and intonation trajectories were extracted. A classifier ensemble model was trained for stress detection, machine learning models were used for rhythm feature analysis, and a convolutional neural network was developed for intonation pattern recognition. The system further integrates diagnostic results into visual feedback, comparative listening, and graded practice modules. The results show that the proposed models can effectively identify learners’ prosodic errors across stress, rhythm, and intonation dimensions. The teaching experiment indicates that the system significantly improves learners’ prosody perception and pronunciation naturalness, providing a data-driven technical route for intelligent and precise oral pronunciation guidance.