A Pronunciation Practice System Based on Pre-trained Deep Learning Models

Advancements in speech-to-text technology have made Automatic Speech Recognition (ASR) an increasingly valuable tool in language learning, easing the burden on teachers who traditionally assess pronunciation manually. While various deep neural network methods have been explored for pronunciation assessment, models like LSTM and BP neural networks have demonstrated superior performance. In this paper, we propose a novel framework for pronunciation assessment that leverages pre-trained models for both speech and natural language processing to generate a comprehensive score for each word. Experimental results show that our scoring model achieves state-of-the-art performance across different network architectures. Additionally, we have deployed the pronunciation practice system on a web platform, allowing users to experience the system firsthand. The system is available at https://117.0.36.6:5000/ .