In order to meet the requirements of voicieprint recognition in combustible and explosive environments, a linear Sagnac interference optical fiber sound sensing system is constructed. Wiener filter algorithm is used to reduce noise of speech data, pitch period features are obtained by three-level clipping method, speaker samples are screened by dynamic time regularizing algorithm, and features of Mel frequency cepstrum coefficient are extracted. An experimental study on voiceprint recognition is carried out using Gaussian mixture model-expectation maximization algorithm. The experimental results show that the sensor system can accurately detect 400 text-independent speech segments between 3s and 5s on the 10km sensor fiber and 2m away from the sound source, and the comprehensive recognition accuracy rate is 92.25%, which is expected to be applied to the field of voiceprint detection and recognition such as equipment failure and emergency rescue in flammable and explosive environments.