Artificial Bandwidth Extension of Narrowband Speech Signals for the Improvement of Perceptual Speech Communication Quality

In this paper, an artificial bandwidth extension (ABE) algorithm from narrowband to wideband is proposed in order to improve the quality of narrowband speech. The proposed ABE algorithm is based on spectral band replication in the modified discrete cosine transform (MDCT) domain with no additional bits. In particular, the patch index search for the replication band is restricted so that the harmonic structure of the wideband speech is maintained after ABE. In the proposed ABE algorithm, we first determine whether the current analysis frame of speech signals is voiced or unvoiced. A harmonic spectral replication or a correlation-based replication approach is then applied for the voiced or unvoiced frame, respectively. The proposed ABE algorithm is finally embedded into the G.729 speech decoder as a post-processor. It is shown from the subjective evaluation using a MUSHRA test that the mean opinion score of the wideband speech signals extended by the proposed ABE method is measured as 75.5, which is higher of around 14% than that of narrowband speech signals.