Optical fiber speech sensors have the advantages of anti-electromagnetic interference, high sensitivity, and long sensing distance. However, there are often much noise when using them for speech detection, which greatly affects the quality of speech signals. To reduce the influence of noise on the speech signal, a multiwindow spectrum estimation spectrum subtraction method based on endpoint detection for the optical fiber speech sensor system is proposed in this paper. Through endpoint detection, it is determined whether the signal is a speech frame to realize the estimation of the average noise amplitude. The constructed linear Sagnac optical fiber speech sensing experiment results show that compared with traditional spectral subtraction and multiwindow spectrum estimation spectrum subtraction methods, the method has a better suppression effect on background noise, and the signal-to-noise ratio of the speech signal can be increased by 2?3 dB.