Synthesis of enhanced audio from low bitrate compressed audio based on unit selection and statistical conversion methods

Audio synthesis algorithms in general, are aimed at synthesizing audio based on target specifications or target descriptors with the requirement of natural-sounding results and correct content representation. However, work on synthesis algorithms in an audio quality enhancement context, is limited. In this paper, a new algorithm on audio synthesis is presented that attempts to improve the quality of compressed audio under the constraint that no specific information on the original, uncompressed signal is available. The methods employed here combine corpus-based audio synthesis techniques and the statistical spectral conversion framework adopted from speech conversion algorithms. The results show significant enhancement of a compressed audio signal under subband-specific and audio quality evaluation tests.