A study of HMM-based bandwidth extension of speech signals

In this paper, we investigate a representative statistical method for artificial bandwidth extension: the hidden Markov model (HMM) based method. In particular, we are interested in objectively quantifying its performance using both static and dynamic measures. The Gaussian mixture model (GMM) based method is presented as a reference method for the performance test under various HMM configurations. We also reasonably claim that a general approach using Baum–Welch re-estimation algorithm performs better than the existing training algorithm suggested by Jax. Accordingly, it is used as the basic algorithm for the training of HMM model. Test results show that the static performance of HMM-based method depends only on the total number of Gaussian components of HMM model, while its dynamic performance depends dominantly on the number of states of the model. More specifically, it is also observed that the GMM-based method is quite comparable with the HMM-based one in static performance, but, in dynamic performance, the latter outperforms the former even with higher computational complexities.