To provide a speech separation method and a device, capable of accurately separating a speech signal including a non-linear component, while suppressing the number of microphones used for collecting sound to minimum requirement.

The sound signal which is input via two microphones installed in a vehicle is sampled, and sampled data are stored for M pieces at a time, and when speech uttered by a driver is included in the stored data (hereinafter refers to as input signals) x1(1) to x1(M), and x2(1) to x2(M) (S110 to S230), linearizing processing for linearizing the input signals x1 and x2 which contain nonlinear components is performed (S140), and separation processing for separating the signals into a signal component and a noise component is performed, on the basis of the linearized signals z1 and z2 (S150). Then, driver speech signals S(1) to S(M) which are separated by the separation processing are supplied to a speech recognition processing section 13 (S160).