The environment for training and recognition in Chinese speech recognition under computer-aided design may vary due to the difference of channel and background noise. When the trained model cannot well represent the test data, the recognition rate of the system will drop sharply. The computer-aided design method focuses on using a small amount of Chinese voice data to improve the performance of the system in the test environment. In this paper, we choose the BiLSTM CRF word separation model under deep learning as the improved benchmark model, and combine the Bert language pre-training module to enhance the performance of Chinese word separation. Combining the deep learning sample transfer learning theory and the improved sampling strategy, an adaptive translation model for intelligent Chinese domain is constructed. The experimental results show that Bert Chinese word segmentation model is superior to other word segmentation models in different data sets and has the best word segmentation performance, which can provide reliable support for the application experiment of this model. The test results show that this method can achieve high speech recognition accuracy and good application results.