Traditional prediction-error techniques for multivariable system identification require canonical descriptions using a large number of parameters. This problem can be avoided using subspace based methods, since these estimate a state-space model directly from the data. The main computations consist of a QR-decomposition and a singular-value decomposition. Herein, a subspace based technique for identifying general finite-dimensional linear systems is presented and analyzed. The technique applies to general noise covariance structures. Explicit formulas for the asymptotic pole estimation error variances are given. The proposed method is found to perform comparable to a prediction error method in a simple example.