The purpose of speech recognition technology is to transform human voice into computer-understandable sound signals, which is a key technical aspect of human-computer interaction. In this paper, we propose an improvement method of fundamental cycle extraction method, use wavelet decomposition and high-order cumulative volume to realize reconstruction and analysis of low-frequency signals, and put forward a convolutional neural network of one-dimensional vectors to simplify the computation procedure, based on which we realize the objective evaluation of timbre accuracy and pitch control. In the example analysis of timbre accuracy, the algorithm in this paper is closer to the evaluation results of professional vocal teachers. In the evaluation of pitch control, the average error of this algorithm for two students is 1.824% and 2.681%. After using the algorithm, the number of students with high scores increased by 45.5%, 85% of the students believed that their interest in vocal lessons had increased significantly, and 81% of the students believed that their vocal skills had increased significantly after using the algorithmic evaluation tool.