With the development of Internet of Things technology, the application of intelligent speech recognition in educational platforms has become increasingly widespread. Traditional noise suppression techniques lack flexibility in dealing with noise with both high and low frequencies, and cannot maintain high accuracy in various noise scenarios. To solve this problem, this article collected speech and visual signals through the fourth generation Amazon Echo Dot, used a Multimodal Generative Adversarial Network (MGAN) to generate simulated speech signals under noise, and combined the U-Net model for noise suppression. Furthermore, the Wav2Vec 2.0 model was used to capture contextual information in speech signals, thereby achieving efficient and accurate speech recognition in a multi-noise environment. The experimental results indicate that the proposed method obtains a speech recognition accuracy of 94.3% in noisy environments, significantly better than existing traditional methods, effectively improving the speech interaction effect in the Russian MOOC (Massive Open Online Course) system and enhancing the user learning experience.