Thanks to the powerful representation and generation capability of language models, grammatical error correction models have achieved outstanding performances. However, the grammatical error correction community still faces the over-correction challenge without external linguistic knowledge injection. To address this issue, we propose a novel dynamic linguistic knowledge fused language model for grammatical error correction. In this work, we regard Chinese-specific pronunciation representation “pinyin” and detection information as extra-linguistic knowledge to improve the correction accuracy in both traditional Sequence-to-Sequence (Seq2Seq) model and large language models. For the Seq2Seq model, we design an adaptive network to integrate pinyin into the pre-trained language model to boost its linguistic understanding ability. Meanwhile, we exploit error detection as an auxiliary task to ensure the shared encoder implies potential detection information. For large language models, we exploit large language models to self-extract linguistic knowledge and integrate them into prompt instructions to further enhance its language understanding capabilities. Experiments on several benchmark datasets show that our model can consistently improve grammatical error correction performance, leading to state-of-the-art results on all datasets. Detailed analysis gains more insights into the contribution of pinyin and detection knowledge.