Magnetic resonance imaging (MRI) is crucial for medical diagnosis. However, the loud noise produced by MRI devices prevents doctors from hearing the patient's voice. In a previous study, voice detection employing deep learning exhibited low performance for the various types of MRI noise. The various types of noise generated by different MRI techniques became an issue in that the voice detection accuracy decreased in the presence of MRI noise. In the present work, we developed an approach to detect the patient's voice with high accuracy when mixed with the various types of MRI noise. We propose an approach to enhance the voice detection accuracy for various types of MRI noise by training the voice detection model on data generated by a variational autoencoder (VAE). In this approach, the VAE learns new types of MRI noise while remembering previous MRI noise by continuous learning. This is a new approach for assessing continuous learning in loud environments with various types of noise, including MRI noise. Experiments were performed to examine the effectiveness of the voice detection system with VAE. The average F-measure for three types of MRI noise was improved by 0.008, and the F-measure for previous training data was improved by up to 0.247 with respect to the conventional method. Thus, this approach was more effective for training long-term memory, and can be used for voice detection in the presence of various types of MRI noise by repeatedly learning new MRI noise.