How does one calculate a steering matrix for DOA given real data and only antenna positions

I am currently trying to implement the MUSIC algorithm for a real dataset which captures a sound signal in a room using a cubic 8 sensor array setup.

Every example I've found for the MUSIC algorithm talks about the steering matrix and calculates it using a set of DOAs which are, well, the actual DOAs that the algorithm is trying to estimate? This part deeply confuses me.

All I have as inputs are: X = [M x N] matrix, where M is the number of antennas, N is the number of timesteps of the signals. So essentially an 8 channel signal snapshot. Y = [M x 3] matrix, which holds the 3D coordinates of each sensor.

It is worth noting that the sensors in my dataset are not actually a perfect cuboid. It's essentially two rectangles where the upper rectangle is above the other, but also slightly shifted in the X axis.

It is my understanding that a steering matrix should be computed using only Y. X is used later, for the eigenvector computation. So how would I go about generating this matrix if all I have are mic positions?

Any advice, literature would help me greatly.