To reduce the noise floor in interferometric fiber-optic sensing systems based on 3 × 3 coupler demodulation, this paper proposes spectral subtraction based on multi-window minimum statistical noise estimation to process the two interference signals. This method estimates the noise spectrum of three types of windowed signals, averages these estimation results, and subtracts them from the signal, effectively removing noise from the interfering signal before demodulation, thereby minimizing noise aliasing during demodulation. Subsequently, the ellipse fitting algorithm and the arctangent algorithm are utilized to accomplish signal demodulation. The experimental results indicate that the system's noise floor is reduced by an average of 25 dB within the 0-5 kHz frequency range and achieves a maximum reduction of 30 dB to a level of 0.7 µrad/√Hz at 50 kHz. This scheme does not use additional optics devices, reducing the complexity and cost of the system, and thereby enhancing the feasibility of interferometric fiber optic sensing systems in various environments.