Depth of weld penetration is very important to laser welding quality. Laser welding is a complicated process, and quantitative analysis of this process is quite difficult. A set of TC4 titanium alloy thin plate specimens are used as laboratory samples. The acoustic signals are first preprocessed by the spectral subtraction noise reduction method and analyzed in both time and frequency domains, and a valid relationship between the acoustic signals and the weld penetration depth is deduced. Radial basis function neural network (RBFNN) models are developed to predict the weld penetration depth. Sound pressure deviation, band power, laser power and welding speed are used as input variables of RBFNN. The results show that the acoustic signal can characterize and predict the depth of weld penetration well under different laser welding parameters.