This paper describes a new algorithm for computing linear generators (vector generating polynomials) for matrix sequences, running in sub-quadratic time. This algorithm applies in particular to the sequential stage of Coppersmith's block Wiedemann algorithm. Experiments showed that our method can be substituted in place of the quadratic one proposed by Coppersmith, yielding important speedups even for realistic matrix sizes. The base fields we were interested in were finite fields of large characteristic. As an example, we have been able to compute a linear generator for a sequence of 4*4 matrices of length 242 304 defined over GF(2^607) in less than two days on one 667MHz alpha ev67 cpu.