Automatic Speech Recognition (ASR) has been widely used in intelligent customer service, voice assistants, and in-vehicle systems, but its output often contains grammatical errors, near-sound word substitution, and other problems. These problems also exist in the resource-poor Uyghur language, reducing the readability of text and affecting downstream tasks such as machine translation and text summarization. To alleviate this problem, we propose an adaptive multi-source feature hierarchical fusion model for text error correction. In order to improve the robustness of the model in dealing with such errors, and taking into account the typical error characteristics of Uyghur ASR texts, a dedicated corpus for text error correction task training covering both phonemic and syllabic information is firstly constructed, which is used to enhance the model’s ability to perceive the pronunciation features and the morphological structure inside the word. On this basis, the proposed error correction model introduces three types of representations, namely text, phoneme and syllable, in parallel at the encoding stage, and effectively integrates semantic and pronunciation information through an adaptive hierarchical fusion strategy. Experimental results demonstrate that the proposed framework reduces the word error rate by approximately 10% compared to specialized text error correction frameworks, while maintaining stable performance on public English datasets, demonstrating good generalizability.