In stochasticoding, the speech waveform is represented as a nonstationary Gaussian random process and is reconstructed by filtering an i.i.d. Gaussian source (innovation) with a time‐varying linear filter [M. R. Schroeder and B. S. Atal, Proc. ICASSP 85]. The design of optimum block codes for the white Gaussian innovation using a “polar” product code representation is considered. Each codeword in the product codebook consists of a gain and a vector's “orientation” (specified by a set of vectors on the unit hypersphere). The amplitude factors of the codewords are encoded with a Max quantizer. The codeword “orientations” are iteratively generated from an ensemble of random spherical vectors so as to maximize the minimum spherical angle between the orientation vectors. The product codes are jointly searched to determine the optimum Gaussian innovation. This approach is compared with probabilistically generated codes as well as iteratively generated (cluster) codes for performance. Preliminary simulation results have shown an improvement of nearly 1 dB in the average SNR.