This paper presents two versions of an algorithm for bandwidth extension of speech signals. We focus on the generation of the spectral envelope and compare the performance of two different approaches - neural networks versus codebooks - in terms of objective and subjective distortion measures.