<p>Due to the assumption of conditional independence, the connectionist temporal classification (CTC) model usually needs a language model to improve its speech recognition performance, which will increase complexity and computing costs. Therefore, a simple and effective method named information-preserving multilayer CTC loss (IP-MLCTC) is proposed in this paper. Unlike the traditional CTC model that only calculates the CTC loss of its last layer, CTC losses of different layers are calculated to get a multilayer CTC loss. This multilayer CTC loss comprehensively considers the information of different layers. Minimizing it will maximize the contribution of each layer. In addition to utilizing information from different layers for loss calculation during training, we fuse the information from different layers instead of using that from only the last layer during the inference. This can prevent information loss and apply the contribution of each layer’s information to inference. It ensures consistency between training and inference and further improves performance. The AISHELL-1 was used for training and inference, while the multi-scenario datasets published by SpeedColab were used for inference. The results demonstrate that the proposed method can relatively reduce the character error rate by 7.7% without adding a language model, thus achieving better speech recognition performance.</p>