With the deepening of economic and cultural globalization and the popularity of cross-cultural communication, Mandarin, as a key carrier of Chinese culture, has become increasingly important for both domestic language education and foreign Chinese learning. However, traditional Mandarin teaching faces limitations such as difficulty in real-time detection of individual reading errors (e.g., missing reading and back reading) and heavy reliance on teacher experience, which restricts the efficiency of error correction and teaching quality. Meanwhile, with the rapid development of information technology, deep learning has shown strong advantages in speech signal processing, providing a new technical path for intelligent Mandarin reading error detection. Against this background, this study focuses on the detection of missing reading and back reading errors in Mandarin reading aloud, and conducts research based on the deep learning framework. To improve the accuracy and efficiency of error detection, this study takes the traditional Deep Neural Network (DNN) as the basic model, and optimizes the core Reading Quality Assessment (GOP) algorithm: first, it extends the GOP algorithm to the DNNbased error detection system, and modifies the GOP calculation formula by introducing the average posterior probability of non-target senones and weight coefficients, which solves the problem of unreliable phoneme segmentation caused by nonstandard pronunciation; second, it addresses the issue that missing-reading errors of the current phoneme affect the GOP calculation of adjacent phonemes in the traditional framework, further optimizing the algorithm’s robustness. Additionally, this study introduces DNN adaptive technology based on KL divergence regularization to align the standard and nonstandard reading models, enhancing the algorithm’s adaptability to different speakers. Experiments are conducted on two databases (MPE database for domestic Mandarin speakers and ICALL database for foreign Chinese learners). The results show that the improved GOP algorithm combined with DNN adaptive technology significantly outperforms traditional methods: compared with the GMM-CM algorithm, the accuracy and recall of error detection are increased by 13.4%; compared with the original DNN-GOP algorithm, the improved DNN-GOP2 algorithm reduces the Top1 error rate by 1.7% and the Top5 error rate by 2.0%. This study not only provides a more accurate and efficient technical solution for Mandarin reading error detection but also lays a foundation for the development of intelligent Mandarin teaching systems, which is of great significance for promoting the modernization of Mandarin teaching and the popularization of Chinese language education globally.