The current speech recognition system still has limited accuracy when faced with a variety of accents, noisy environments, and changes in speech speed. Especially for non-native speakers, accents, non-standard pronunciation, and mixed languages can easily lead to model recognition errors, which in turn affect the accuracy of scoring. This paper introduces a pronunciation assessment model based on convolutional neural networks (CNN) for automatic pronunciation scoring on an English speaking training platform. First, this paper collects audio data through a high-quality microphone and preprocesses it, including noise removal, normalization, and audio frame segmentation. Then, the Mel-frequency cepstral coefficients (MFCC) and other speech features (spectral centroid, zero-crossing rate, pitch, etc.) are used to extract audio features to provide input data for model training. In terms of model construction, this paper designs a CNN-based deep learning architecture, which extracts local features of audio signals through multi-layer convolution and pooling operations, and scores pronunciation through a fully connected layer. In order to evaluate the performance of the model, the experiment divides the dataset into a training set and a validation set and uses the cross entropy loss function and the Adam optimizer for training. During the training process, the model’s accuracy, recall, precision, and F1 score performed well (above 0.7), verifying the effectiveness of CNN in the pronunciation evaluation task.