To address the problem that traditional artificial recognition of oral English in the development of oral English is easy to be affected by its own factors leading to the recognition accuracy is not high, this article adopted the long short-term memory (LSTM) neural network in machine learning algorithms to construct a speech recognition model. Pre-emphasis, framing, windowing, Mel Frequency Cepstral Coefficient (MCFF) feature extraction and other preprocessing operations were performed on the speech data. Using MFCC parameters as model inputs to analyze and validate the performance of speech recognition models, multiple parameters such as intonation, speaking speed, and rhythm were utilized to evaluate oral speech. The findings showed that the model constructed in this article achieved a speech recognition accuracy of 95.21%, which can effectively enhance the accuracy of speech recognition.