Laser Doppler vibrometry-based voice detection systems can measure voice signals without direct contact by demodulating the Doppler signals caused by sound vibrations. However, in remote voice detection, the coherent signal is significantly affected by noise in the detection environment. This weakens the energy of the signal and reduces the signal-to-noise-ratio (SNR), resulting in the signal being incorrectly demodulated. We present a coherent signal demodulation method based on generative adversarial networks (GANs) called a phase demodulation generative adversarial network (PDGAN) that can overcome this problem. To develop the method, coherent signal and Doppler signal datasets were also established. Demodulation from a coherent signal is accomplished through unsupervised learning within the PDGAN, layer by layer, with global supervised feedback learning for fine-tuning. Experiments confirm that the PDGAN method can demodulate Doppler signals when the SNR is as low as −20  dB, with the demodulation effect being notably better than traditional phase demodulation methods. The PDGAN method is also robust against noise, so it can extend the measurement distance for laser voice detection. Generally, the method’s performance confirms the applicability of deep learning approaches for optical measurement.