The laser Doppler velocimeter (LDV) is commonly used in high-G accelerometer shock testing to provide high-precision reference velocity measurements. However, noise inevitably interferes with LDV signals, reducing the measurement accuracy. A novel denoising method based on convolutional neural network with long short-term memory (CNN-LSTM) and complementary ensemble empirical mode decomposition (CEEMD) is proposed to improve the measurement accuracy of reference velocity. First, the weights were obtained by training the constructed CNN-LSTM neural network. CEEMD was then used to process the training signals, and the resulting IMF was partially zeroed. Furthermore, the splitting points were evaluated and optimized. Finally, the weights and optimal splitting points were applied to the test signals. Simulation and experimental results show that the proposed method outperforms wavelet thresholding and CNN-LSTM in denoising performance. The results show that the proposed method can improve the accuracy of the demodulated velocity and thus contribute to accelerometer shock testing.