Vector quantisation (VQ) is a method widely used in low bit-rate coding and transmission of speech signals. Unfortunately, a single bit error in the transmitted index, due to noise in the transmission channel, could degrade perceived speech quality at the receiver quite dramatically, as the reference vector retrieved by the corrupted index may di er greatly from the vector corresponding to the intended index. The index assignment (IA) process (an NP-complete combinatorial optimisation problem) attempts to re-order the code book to minimise the e ects of single-bit errors, but generally only at considerable computational expense. This paper presents an improved vector quantisation algorithm, based on Kohonen's Self-Organising Feature Map (K-SOFM), that jointly optimises the quantisation error and resistance to noise in the transmission channel. This is achieved using a neighbourhood function based on the Hamming distance between code book indices, rather than the normal Euclidean distance across a two dimensional feature map. As a result, similar reference vectors are recalled by indices with similar binary patterns, minimising the e ect of errors in the transmitted index introduced by noise in the transmission channel.