This study constructs an interactive teaching platform based on speech recognition technology for language courses in colleges and universities. The study begins with preprocessing speech signals, including sampling, enhancement, pre-emphasis and other steps. Subsequently, speech features are extracted using Improved Mel Falling Spectral Coefficients (IMFCC), and features are optimized by combining Principal Component Analysis (PCA). In addition, the study uses the Long Short-Term Memory (LSTM) network and Gated Recurrent Unit (GRU) to improve speech recognition’s accuracy and efficiency. The experimental results show that the method in this study achieves an average accuracy of 98% in classroom speech emotion recognition, which is a significant improvement over the traditional ADASYN method. This paper also provides an in-depth study of classroom interaction behavior by comprehensively analyzing the speech recognition results. The study shows that the platform can effectively promote the interaction between teachers and students and improve the quality of teaching. This study provides an efficient interactive tool for language teaching in colleges and universities, which helps to realize a personalized and dynamic teaching process.